feat(chats): Enable web search via Vertex Gemini model #751
Conversation
junaid-shirur
requested review from
kalpadhwaryu,
shivamashtikar and
zereraz
as code owners
|
Note Other AI code review bot(s) detectedCodeRabbit has detected other AI code review bot(s) in this pull request and will avoid duplicating their findings in the review comments. This may lead to a less comprehensive review. WalkthroughAdds optional Web Search across frontend and server: an enableWebSearch flag flows through routes, hooks, and APIs; AI providers (Gemini/Vertex) and types support web-search tooling, sources, and grounding; chat streaming and citation SSEs surface web sources; search types/icons/UI updated; vespa file-ID queries extended. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
actor User
participant Frontend as Frontend Chat UI
participant Router as Chat Route
participant Hook as useChatStream
participant API as /api/v1/message/create
participant Server as server/api/chat
participant Provider as AI Provider (Gemini/Vertex)
User->>Frontend: Submit question (enableWebSearch?)
Frontend->>Router: Navigate with query params (enableWebSearch)
Router->>Hook: startStream(..., enableWebSearch)
Hook->>API: SSE request ?enableWebSearch=true|false
API->>Server: route handling (enableWebSearch)
Server->>Provider: converseStream({ webSearch: flag })
alt Web Search enabled
Provider->>Provider: enable web tools + grounding
Provider-->>Server: chunks + sources + groundingSupports
Server->>Frontend: SSE chunks + CitationsUpdate events
else Disabled
Provider-->>Server: chunks (text only)
Server-->>Frontend: SSE chunks
end
Frontend-->>User: Render response (+ citations)
sequenceDiagram
autonumber
participant Index as server/ai/provider/index.ts
participant Vertex as VertexAiProvider
note over Index,Vertex: Dynamic Vertex backend selection
Index->>Vertex: new VertexAiProvider({projectId,region,provider})
alt provider = GOOGLE
Vertex->>Vertex: instantiate Google VertexAI client
Vertex->>Vertex: use Google-specific converse/converseStream (webSearch tooling)
else provider = ANTHROPIC
Vertex->>Vertex: instantiate AnthropicVertex client
Vertex->>Vertex: use Anthropic-specific converse/converseStream
end
Index->>Vertex: webSearchQuestion(..., webSearch=true)
Vertex-->>Index: stream text + sources + groundingSupports
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related PRs
Suggested reviewers
Poem
✨ Finishing Touches
🧪 Generate unit tests
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
SupportNeed help? Create a ticket on our support page for assistance with any issues or questions. CodeRabbit Commands (Invoked using PR/Issue comments)Type Other keywords and placeholders
CodeRabbit Configuration File (
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @junaid-shirur, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request introduces a significant enhancement to the chat functionality by integrating real-time web search capabilities. The primary goal is to improve the accuracy and recency of AI responses by allowing the model to access up-to-date information from the web. The changes involve updates to how chat messages are processed, how AI models are configured, and how search results are presented to the user.
Highlights
- Web Search Integration: Introduced web search capabilities leveraging the Vertex Gemini model, allowing the AI to fetch real-time information to answer user queries. This includes adding an
enableWebSearchparameter across various frontend and backend functions and API calls. - Frontend UI Enhancements: Updated the chat interface to support web search. This involves adding a new
WebSearchEntityandApps.WebSearchfor proper display of web search sources and a dedicated globe icon to represent web search results. - Backend AI Provider Updates: Modified
GeminiAIProviderandVertexAiProviderto incorporategoogleSearchtools within their model configurations when web search is enabled. TheVertexAiProvidernow supports both Anthropic and Google Vertex backends, with specific implementations for each. - Advanced Citation Handling: Implemented new logic in the backend to process and integrate web search citations and grounding supports from Gemini/Vertex AI responses directly into the chat UI, providing users with verifiable sources for AI-generated answers.
- Type Definition Extensions: Extended core type definitions (
ModelParams,ConverseResponse) to include new properties likewebSearch,sources, andgroundingSupports, ensuring robust data handling for the new web search feature.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a web search feature powered by the Vertex Gemini model. The changes span both the frontend and backend, adding a toggle for web search in the UI and routing requests to the appropriate model on the backend. The VertexAiProvider has been significantly refactored to support multiple underlying providers (Anthropic and Google), which is a great improvement for flexibility. My feedback includes a critical fix for a logic issue in the frontend routing parameters, along with several suggestions to improve type safety and performance on the backend.
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 8
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (5)
server/search/vespa.ts (1)
846-899: Escape YQL inputs and dedupe IDs to prevent injection and reduce bloat
- collectionIds/collectionFolderIds/collectionFileIds are interpolated directly into YQL without escaping; this allows quote-breaking and potential YQL injection if any ID contains
'. Use the existing escapeYqlValue helper.- Also dedupe/trim IDs to avoid redundant OR clauses.
Apply:
- const buildCollectionFileYQL = async () => { + const buildCollectionFileYQL = async () => { // Extract all IDs from the key-value pairs - const collectionIds: string[] = [] - const collectionFolderIds: string[] = [] - const collectionFileIds: string[] = [] + const collectionIds: string[] = [] + const collectionFolderIds: string[] = [] + const collectionFileIds: string[] = [] for (const selection of collectionSelections) { if (selection.collectionIds) { collectionIds.push(...selection.collectionIds) } if (selection.collectionFolderIds) { collectionFolderIds.push(...selection.collectionFolderIds) } if (selection.collectionFileIds) { collectionFileIds.push(...selection.collectionFileIds) } } - let conditions: string[] = [] + // Normalize and dedupe + const uniq = (arr: string[]) => + Array.from(new Set(arr.filter(Boolean).map((s) => s.trim()))) + const normCollectionIds = uniq(collectionIds) + const normFolderIds = uniq(collectionFolderIds) + const normFileIds = uniq(collectionFileIds) + const conditions: string[] = [] // Handle entire collections - use clId filter (efficient) - if (collectionIds.length > 0) { - const collectionCondition = `(${collectionIds.map((id: string) => `clId contains '${id.trim()}'`).join(" or ")})` + if (normCollectionIds.length > 0) { + const collectionCondition = `(${normCollectionIds + .map((id: string) => `clId contains '${escapeYqlValue(id)}'`) + .join(" or ")})` conditions.push(collectionCondition) } // Handle specific folders - need to get file IDs (less efficient but necessary) - if (collectionFolderIds.length > 0) { - const clFileIds = await getAllFolderItems(collectionFolderIds, db) + if (normFolderIds.length > 0) { + const clFileIds = await getAllFolderItems(normFolderIds, db) if (clFileIds.length > 0) { const ids = await getCollectionFilesVespaIds(clFileIds, db) - const clVespaIds = ids - .filter((item: any) => item.vespaDocId !== null) - .map((item: any) => item.vespaDocId!) + const clVespaIds = Array.from( + new Set( + ids + .filter((item: any) => item.vespaDocId != null) + .map((item: any) => String(item.vespaDocId)) + ) + ) if (clVespaIds.length > 0) { - const folderCondition = `(${clVespaIds.map((id: string) => `docId contains '${id.trim()}'`).join(" or ")})` + const folderCondition = `(${clVespaIds + .map((id: string) => `docId contains '${escapeYqlValue(id)}'`) + .join(" or ")})` conditions.push(folderCondition) } } } // Handle specific files - use file IDs directly (most efficient for individual files) - if (collectionFileIds.length > 0) { - const ids = await getCollectionFilesVespaIds(collectionFileIds, db) - const clVespaIds = ids - .filter((item: any) => item.vespaDocId !== null) - .map((item: any) => item.vespaDocId!) + if (normFileIds.length > 0) { + const ids = await getCollectionFilesVespaIds(normFileIds, db) + const clVespaIds = Array.from( + new Set( + ids + .filter((item: any) => item.vespaDocId != null) + .map((item: any) => String(item.vespaDocId)) + ) + ) if (clVespaIds.length > 0) { - const fileCondition = `(${clVespaIds.map((id: string) => `docId contains '${id.trim()}'`).join(" or ")})` + const fileCondition = `(${clVespaIds + .map((id: string) => `docId contains '${escapeYqlValue(id)}'`) + .join(" or ")})` conditions.push(fileCondition) } }Optional follow-ups:
- If selections can be very large, chunk OR clauses (e.g., 500 IDs per group) to avoid query-size limits.
- Consider try/catch around DB lookups to degrade gracefully (log and skip this subquery) instead of failing the whole agent search.
frontend/src/routes/_authenticated/index.tsx (1)
223-241: FixenableWebSearchpropagation, schema, and ChatBox support
In
frontend/src/routes/_authenticated/index.tsx(around lines 223–241, 247–253, 267–270), change- if (enableWebSearch) { - searchParams.enableWebSearch = enableWebSearch - } + if (enableWebSearch !== undefined) { + searchParams.enableWebSearch = enableWebSearch + }so
falsevalues are included.In
frontend/src/routes/_authenticated/chat.tsx(lines 2793–2801), replace the string‐based schemaenableWebSearch: z .string() .transform(val => val === "false") .optional() .default("false")with a boolean schema, e.g.:
enableWebSearch: z.boolean().optional().default(false)In
frontend/src/components/ChatBox.tsx(props at ~lines 132–138 and call at ~1729–1732), extendhandleSendto include theenableWebSearchparameter and forward it inhandleSendMessage().frontend/src/routes/_authenticated/chat.tsx (1)
2518-2526: Sanitize after JSON-to-HTML conversion to avoid breaking JSON.Currently you sanitize the JSON string before
JSON.parse, which can corrupt JSON payloads. Build the HTML first, then sanitize that HTML.- __html: jsonToHtmlMessage(DOMPurify.sanitize(message)), + __html: DOMPurify.sanitize(jsonToHtmlMessage(message)),server/ai/provider/vertex_ai.ts (2)
29-66: Image part shape is Anthropic-style; Google Vertex AI expects inlineData parts.
buildVertexAIImageParts()returns{ type: "image", source: { ... } }, which fits Anthropic but not Vertex AI. In the Google paths you pass these parts tosendMessage/sendMessageStream, which expect Vertex AI “Part” objects, e.g.{ inlineData: { mimeType, data } }. This will break image-grounded prompts in the Google provider.
- Split builders: one for Anthropic (current shape), one for Google Vertex AI.
- Use the correct builder in each path.
-const buildVertexAIImageParts = async (imagePaths: string[]) => { +const buildAnthropicImageParts = async (imagePaths: string[]) => { // ... current implementation unchanged ... } + +const buildGoogleVertexImageParts = async (imagePaths: string[]) => { + const baseDir = path.resolve(process.env.IMAGE_DIR || "downloads/xyne_images_db") + const imagePromises = imagePaths.map(async (imgPath) => { + const match = imgPath.match(/^(.+)_([0-9]+)$/) + if (!match) throw new Error(`Invalid image path: ${imgPath}`) + const docId = match[1] + const imageDir = path.join(baseDir, docId) + const absolutePath = findImageByName(imageDir, match[2]) + const ext = path.extname(absolutePath).toLowerCase() + const mimeMap: Record<string, string> = { ".png": "image/png", ".jpg": "image/jpeg", ".jpeg": "image/jpeg", ".webp": "image/webp" } + const mimeType = mimeMap[ext] + if (!mimeType) return null + try { + await fs.promises.access(absolutePath, fs.constants.F_OK) + const imgBuffer = await fs.promises.readFile(absolutePath) + if (imgBuffer.length > MAX_IMAGE_SIZE_BYTES) return null + const base64 = imgBuffer.toString("base64") + return { inlineData: { mimeType, data: base64 } } + } catch (err) { + Logger.error(`Failed to read image: ${absolutePath}`) + return null + } + }) + const results = await Promise.all(imagePromises) + return results.filter(Boolean) +}Use the appropriate builder:
-// Anthropic paths -const imageParts = params.imageFileNames?.length ? await buildVertexAIImageParts(params.imageFileNames) : [] +const imageParts = params.imageFileNames?.length ? await buildAnthropicImageParts(params.imageFileNames) : [] -// Google paths -const imageParts = params.imageFileNames?.length ? await buildVertexAIImageParts(params.imageFileNames) : [] +const imageParts = params.imageFileNames?.length ? await buildGoogleVertexImageParts(params.imageFileNames) : []Also ensure
createLabeledImageContentcan accept Vertex-style parts (it currently pushes parts verbatim).Also applies to: 209-312, 313-455
121-124: AlignimageFileNamesformat between producers and consumers.
extractImageFileNamesandcreateLabeledImageContentgenerate/expectdocIndex_docId_imageNumber, butbuildVertexAIImageParts(and otherbuild*ImageParts) only parsedocId_imageNumbervia/^(.+)_([0-9]+)$/. Update those loaders to use/^(\d+)_(.+)_(\d+)$/, then setdocId = match[2]andimageNumber = match[3](or strip the leadingdocIndex_before parsing) so images load and label correctly.
🧹 Nitpick comments (16)
server/api/search.ts (1)
169-175: Unify boolean parsing and verify retry parityParsing looks fine. To avoid drift with other flags (e.g., isReasoningEnabled), factor a shared boolean parser and reuse it here. Also, if retries should preserve web search behavior, mirror this flag in messageRetrySchema.
Apply locally within this file:
+const booleanString = z + .string() + .optional() + .transform((val) => !!val && val.toLowerCase() === "true") export const messageSchema = z.object({ ... - enableWebSearch: z - .string() - .optional() - .transform((val) => { - if (!val) return false - return val.toLowerCase() === "true" - }), + enableWebSearch: booleanString, ... })If retries must honor the same switch (optional—confirm server behavior):
export const messageRetrySchema = z.object({ messageId: z.string().min(1), agentId: z.string().optional(), agentic: z.string().optional().default("false"), isReasoningEnabled: z .string() .optional() .transform((val) => { if (!val) return false return val.toLowerCase() === "true" }), + enableWebSearch: booleanString.optional(), })frontend/src/hooks/useChatStream.ts (2)
181-181: Reasoning text growth nitAppending “\n” per step improves readability but can balloon memory for long runs. Consider bounding length or storing steps as an array and joining on render.
283-284: Only send enableWebSearch when trueSaves bytes and keeps defaults server-side.
-url.searchParams.append("enableWebSearch", enableWebSearch.toString()) +if (enableWebSearch) { + url.searchParams.append("enableWebSearch", "true") +}server/ai/types.ts (2)
83-84: Name consistency: webSearch vs enableWebSearchRequest/URL uses enableWebSearch; ModelParams uses webSearch. Align names or document the mapping at the API boundary to reduce confusion.
95-109: WebSearchSource/GroundingSupport: clarify index semanticsConsider documenting whether segment indices are UTF-16 code units and whether endIndex is exclusive. If client Citation uses url, either map uri → url server-side or rename for consistency.
server/ai/provider/index.ts (1)
281-296: Env-driven Vertex backend init: log enum label and guard invalid env value
- Logging
${provider}prints the numeric enum; log the label for readability.- If VERTEX_PROVIDER is set to an unknown value, you silently fall back to ANTHROPIC. Log a warning to aid ops.
- const provider = - vertexProviderType && VertexProvider[vertexProviderType] - ? VertexProvider[vertexProviderType] - : VertexProvider.ANTHROPIC + const provider = + vertexProviderType && VertexProvider[vertexProviderType] + ? VertexProvider[vertexProviderType] + : VertexProvider.ANTHROPIC + if (!VertexProvider[vertexProviderType]) { + Logger.warn( + `VERTEX_PROVIDER='${vertexProviderType}' not recognized. Falling back to 'ANTHROPIC'.`, + ) + } ... - Logger.info(`Initialized VertexAI provider with ${provider} backend`) + Logger.info( + `Initialized VertexAI provider with ${ + VertexProvider[provider] + } backend`, + )server/ai/provider/gemini.ts (2)
135-141: Type the tools array to avoid implicit any and future regressionsMake the tools’ type explicit to satisfy TS and make schema changes obvious.
- const tools = [] + const tools: Array<{ googleSearch: Record<string, never> }> = [] ... - // Add tools configuration for web search - tools: tools.length > 0 ? tools : undefined, + tools: tools.length > 0 ? tools : undefined,Apply the same typing in converseStream.
- const tools = [] + const tools: Array<{ googleSearch: Record<string, never> }> = []Also applies to: 148-150, 249-255, 262-262
231-396: Consider emitting token usage in metadata for cost/telemetryDownstream accumulates token usage from chunk.metadata.usage; Gemini code doesn’t attach usage, so UI/metrics miss tokens.
- const text = response.text + const text = response.text + // If available from SDK in future, attach usage: + // const usage = response.usage ?? undefined + // const metadata = usage ? { usage } : undefined ... - return { text, cost, sources, groundingSupports } + return { text, cost, sources, groundingSupports /*, metadata*/ }server/api/chat/chat.ts (3)
4647-4655: Token metrics: support both usage shapesVertex provider emits metadata.inputTokens/outputTokens (not metadata.usage). Capture both to keep analytics intact.
- if (chunk.metadata?.usage) { - tokenArr.push({ - inputTokens: chunk.metadata.usage.inputTokens || 0, - outputTokens: chunk.metadata.usage.outputTokens || 0, - }) - } + const usage = + chunk.metadata?.usage ?? + (chunk.metadata && + typeof chunk.metadata.inputTokens === "number" && + typeof chunk.metadata.outputTokens === "number" + ? { + inputTokens: chunk.metadata.inputTokens, + outputTokens: chunk.metadata.outputTokens, + } + : null) + if (usage) tokenArr.push(usage)
4660-4686: Stream the citation brackets delta so UI matches stored answerYou append “[n]” post-stream but never send the delta to the client. Emit one final ResponseUpdate with only the appended part to keep UI and DB in sync.
- if (citationResult) { - answer = citationResult.updatedAnswer + if (citationResult) { + const prev = answer + answer = citationResult.updatedAnswer sourceIndex = citationResult.updatedSourceIndex ... - if (citationResult.newCitations.length > 0) { + if (citationResult.newCitations.length > 0) { citations.push(...citationResult.newCitations) Object.assign(citationMap, citationResult.newCitationMap) stream.writeSSE({ event: ChatSSEvents.CitationsUpdate, data: JSON.stringify({ contextChunks: citations, citationMap: citationMap, }), }) + // Send the brackets delta appended to the text + const delta = answer.slice(prev.length) + if (delta) { + await stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: delta, + }) + } } }
223-223: Type coherence for GroundingSupportYou import GroundingSupport from @google/genai here, but providers surface ai/types.GroundingSupport. Prefer a single internal type (ai/types) throughout to avoid drift.
Would you like me to convert chat.ts to use the internal type and add a narrow adapter at provider boundaries?
Also applies to: 3866-3936
frontend/src/routes/_authenticated/chat.tsx (2)
721-733: Clear enableWebSearch from URL after processing deep-link.You’re clearing other params but not
enableWebSearch; it sticks in the address bar and re-triggers on refresh.router.navigate({ to: "/chat", search: (prev) => ({ ...prev, q: undefined, reasoning: undefined, sources: undefined, agentId: undefined, // Clear agentId from URL after processing toolsList: undefined, // Clear toolsList from URL after processing metadata: undefined, // Clear metadata from URL after processing + enableWebSearch: undefined, // Clear web search toggle }), replace: true, })
2193-2216: Auto-scroll toggling misses “scrolled down but not bottom” case.You only set
userHasScrolledwhen scrolling up. If a user scrolls down slightly and pauses above bottom, new tokens will yank them to bottom. Use “not at bottom” as the criterion.- if (isAtBottom) { - // User is at bottom, allow auto-scroll - setUserHasScrolled(false) - } else if (scrollTop < lastScrollTop.current) { - // User scrolled up, disable auto-scroll - setUserHasScrolled(true) - } + // Disable auto-scroll whenever user is not at bottom; re-enable at bottom + setUserHasScrolled(!isAtBottom)server/ai/provider/vertex_ai.ts (3)
221-279: Non-image code blocks are fine, but image-less branch assumes all blocks have text.
allBlocks.map((block) => ({ text: block.text }))will emit{ text: undefined }for non-text blocks. Guard for"text" in block.- } else { - // otherwise just pass along the raw blocks - messageParts = allBlocks.map((block) => ({ text: block.text })) - } + } else { + // otherwise just pass along the raw blocks (text-only) + messageParts = allBlocks.filter((b: any) => "text" in b).map((b: any) => ({ text: b.text })) + }Also applies to: 368-385
471-474: Fix stray quote and spacing in injected Anthropic text.There’s an extra
"and missing space:"image(s)as". Clean it up.- text: `You may receive image(s)as part of the conversation. If images are attached, treat them as essential context for the user's question.\n\n" - ${userText}`, + text: `You may receive image(s) as part of the conversation. If images are attached, treat them as essential context for the user's question.\n\n${userText}`,
293-307: Streaming: consider emitting a final summary chunk with metadata.Google streaming never yields a final empty chunk with cost/usage like the Anthropic path. For consistency, emit a final chunk (even with empty text) including
metadata/cost.for await (const chunk of result.stream) { // ... existing per-chunk yields ... } + // Final yield to surface accumulated sources/supports without text + yield { + text: "", + cost: 0, + sources: accumulatedSources.length ? accumulatedSources : undefined, + groundingSupports: accumulatedGroundingSupports.length ? accumulatedGroundingSupports : undefined, + metadata: { model: modelParams.modelId, responseTime: Date.now() }, + }Also applies to: 391-449
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (14)
frontend/src/components/GroupFilter.tsx(2 hunks)frontend/src/hooks/useChatStream.ts(5 hunks)frontend/src/lib/common.tsx(3 hunks)frontend/src/routes/_authenticated/chat.tsx(7 hunks)frontend/src/routes/_authenticated/index.tsx(5 hunks)server/ai/provider/gemini.ts(9 hunks)server/ai/provider/index.ts(3 hunks)server/ai/provider/vertex_ai.ts(6 hunks)server/ai/types.ts(1 hunks)server/api/chat/chat.ts(8 hunks)server/api/search.ts(1 hunks)server/search/types.ts(6 hunks)server/search/vespa.ts(2 hunks)server/shared/types.ts(5 hunks)
🧰 Additional context used
🧬 Code graph analysis (9)
server/search/types.ts (1)
server/shared/types.ts (3)
WebSearchEntity(36-36)SystemEntity(33-33)DataSourceEntity(35-35)
server/ai/provider/gemini.ts (1)
server/ai/types.ts (2)
WebSearchSource(95-99)GroundingSupport(101-108)
server/ai/provider/index.ts (3)
server/ai/provider/vertex_ai.ts (1)
VertexAiProvider(67-485)server/ai/types.ts (2)
ModelParams(69-84)ConverseResponse(86-93)server/ai/agentPrompts.ts (1)
agentSearchQueryPrompt(1026-1346)
frontend/src/lib/common.tsx (1)
server/shared/types.ts (3)
Apps(31-31)SystemEntity(33-33)WebSearchEntity(36-36)
frontend/src/components/GroupFilter.tsx (1)
server/shared/types.ts (2)
Apps(31-31)WebSearchEntity(36-36)
frontend/src/routes/_authenticated/chat.tsx (2)
frontend/src/types.ts (1)
ToolsListItem(58-61)frontend/src/components/FollowUpQuestions.tsx (1)
FollowUpQuestions(12-147)
server/ai/provider/vertex_ai.ts (2)
server/ai/types.ts (3)
ModelParams(69-84)ConverseResponse(86-93)WebSearchSource(95-99)server/ai/utils.ts (1)
createLabeledImageContent(1-33)
server/shared/types.ts (1)
server/search/types.ts (2)
KbItemsSchema(25-25)scoredChunk(244-248)
server/api/chat/chat.ts (2)
server/ai/provider/index.ts (3)
webSearchQuestion(1793-1834)generateSearchQueryOrAnswerFromConversation(1457-1495)jsonParseLLMOutput(501-628)server/ai/types.ts (2)
WebSearchSource(95-99)GroundingSupport(101-108)
🪛 GitHub Actions: TypeScript Build Check
server/ai/provider/vertex_ai.ts
[error] 6-6: bunx tsc -b failed with TS2307: Cannot find module '@google-cloud/vertexai' or its corresponding type declarations.
🔇 Additional comments (27)
frontend/src/hooks/useChatStream.ts (3)
238-239: startStream API extended with enableWebSearchSignature change looks good and matches downstream usage.
627-628: Hook surface extendedExposing enableWebSearch here is consistent with startStream. LGTM.
643-644: Confirm retry handling of enableWebSearchThe
/api/v1/message/retrypath and itsmessageRetrySchemadon’t accept or forwardenableWebSearch, so retried messages won’t run web searches. Either addenableWebSearchto the retry URL inuseChatStream.tsand tomessageRetrySchema, or confirm that web search should remain disabled on retries.server/ai/types.ts (1)
91-93: New response fields: ensure SSE wiring to UI typessources and groundingSupports are added here. Verify server-side SSE maps these to front-end Citation/ImageCitation expectations and that citation indices align with rendered text.
server/search/types.ts (7)
63-63: Apps.WebSearch additionEnum extension looks correct.
92-96: Entity validation updatedIncluding WebSearchEntity in isValidEntity is correct.
176-180: KB entity comment-only tweaksNo functional change. Fine.
210-213: Introduce WebSearchEntityGood separation for web search.
216-216: Schema export addedLooks good.
227-227: entitySchema union updatedCorrectly included.
240-240: Entity type union updatedType surface aligned.
frontend/src/components/GroupFilter.tsx (1)
116-118: Add label for Web Search — looks good.Mapping Apps.WebSearch + WebSearchEntity.WebSearch to "Web Search" integrates cleanly with getName().
server/shared/types.ts (5)
36-37: Re-exporting WebSearchEntity from shared/types is correct.This unblocks frontend imports via "shared/types".
265-286: KB file response schema consolidation — validate downstream assumptions.Combining extends and adding chunk/chunkIndex/chunks_summary/relevance/matchfeatures/rankfeatures is fine. Please confirm consumers (API responses, UI mappers) handle these optional fields and the literal type: KbItemsSchema.
482-482: Minor literal style change is benign.status union using double-quoted literals is a no-op for TS. All good.
505-516: Interface line-wrap changes only.No functional/type surface changes detected.
530-546: Interface additions retain compatibility.These keep the enhanced reasoning types coherent. No issues.
frontend/src/routes/_authenticated/index.tsx (1)
155-157: Persisting agentId from URL — OK.Keeps agent selection stable across refresh/navigation.
frontend/src/lib/common.tsx (1)
145-146: Web Search globe icon — good addition.Icon matches the new entity mapping and sizing pattern.
server/ai/provider/gemini.ts (2)
9-15: Type surface alignment with ai/typesImporting WebSearchSource and GroundingSupport from ../types is good and keeps the provider’s output aligned with ConverseResponse.
197-225: Grounding sources and supports extraction looks correctTransforms groundingMetadata into first-class sources and supports and returns them with the response.
server/api/chat/chat.ts (1)
4498-4529: Preserve conversation history in webSearchQuestion
Your call at server/api/chat/chat.ts:4504 passesmessages: llmFormattedMessages, but thewebSearchQuestionprovider currently ignores this and rebuilds history. Verify the implementation appends the base message toparams.messagesso context isn’t lost.frontend/src/routes/_authenticated/chat.tsx (4)
717-718: Good: flag plumbed into send path.Passing
chatParams.enableWebSearchthrough tohandleSendensures deep-linking via URL can toggle web search.
744-751: Good: handleSend and startStream accept web-search flag.The parameter threading is correct and backward-compatible.
Also applies to: 776-786
389-411: VirtualizedMessages integration looks solid.
- Including
currentRespinallItemsprevents layout jumps.measureElement+ absolute positioning is correct with react-virtual.- Keys and memoization look safe.
Also applies to: 2093-2415
641-656: Feedback map extraction tolerant to legacy/new formats.This compatibility layer is tidy and side-effect free.
server/ai/provider/vertex_ai.ts (1)
121-128: Anthropic paths use the right client; token accounting placeholder is acceptable.The flow looks correct; cost accounting can be added later.
Also applies to: 156-165
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (4)
server/api/chat/chat.ts (4)
3866-3873: Nice: helper is now strongly typedprocessWebSearchCitations uses WebSearchSource[] and GroundingSupport[]. Good improvement over any[].
3885-3929: Fix citation insertion index-shift (insert right-to-left or track offset)Inserting at segment.endIndex mutates string length and invalidates subsequent indices. Sort supports by endIndex desc (or track an offset).
- for (const support of finalGroundingSupports) { - const segment = support.segment + // Insert from right to left to avoid shifting downstream indices + const supportsSorted = [...finalGroundingSupports] + .filter( + (s) => + s?.segment && + Number.isFinite((s.segment as any).endIndex as number), + ) + .sort((a, b) => b.segment.endIndex - a.segment.endIndex) + + for (const support of supportsSorted) { + const { endIndex } = support.segment const groundingChunkIndices = support.groundingChunkIndices || [] let citationText = "" for (const chunkIndex of groundingChunkIndices) { if (allSources[chunkIndex]) { const source = allSources[chunkIndex] let citationIndex: number if (urlToIndexMap.has(source.uri)) { // Reuse existing citation index citationIndex = urlToIndexMap.get(source.uri)! } else { citationIndex = sourceIndex const webSearchCitation: Citation = { docId: `websearch_${sourceIndex}`, title: source.title, url: source.uri, app: Apps.WebSearch, entity: WebSearchEntity.WebSearch, } newCitations.push(webSearchCitation) - newCitationMap[sourceIndex] = - citations.length + newCitations.length - 1 - urlToIndexMap.set(source.uri, sourceIndex) + newCitationMap[citationIndex] = + citations.length + newCitations.length - 1 + urlToIndexMap.set(source.uri, citationIndex) sourceIndex++ } citationText += ` [${citationIndex}]` } } - if ( - citationText && - segment?.endIndex !== undefined && - segment.endIndex <= answerWithCitations.length - ) { - answerWithCitations = - answerWithCitations.slice(0, segment.endIndex) + - citationText + - answerWithCitations.slice(segment.endIndex) - } + if (citationText && Number.isFinite(endIndex)) { + const insertAt = Math.min( + Math.max(0, endIndex), + answerWithCitations.length, + ) + answerWithCitations = + answerWithCitations.slice(0, insertAt) + + citationText + + answerWithCitations.slice(insertAt) + } }
4509-4518: Disable “reasoning” for Vertex web-search; rely on per-chunk flagsVertex web-search streams don’t emit Start/EndThinkingToken; leaving reasoning=true routes all tokens to “thinking”.
- searchOrAnswerIterator = webSearchQuestion(message, ctx, { + searchOrAnswerIterator = webSearchQuestion(message, ctx, { modelId: Models.Gemini_2_5_Flash, stream: true, json: false, agentPrompt: agentPromptValue, - reasoning: - userRequestsReasoning && - ragPipelineConfig[RagPipelineStages.AnswerOrSearch].reasoning, + // Let provider chunk.reasoning decide; avoid token-gate markers + reasoning: false, messages: llmFormattedMessages, webSearch: true, })
4594-4630: Route streamed tokens by chunk.reasoning instead of Start/EndThinkingTokenThis branch should honor chunk.reasoning; current gating can swallow output with Vertex.
- if (chunk.text) { - if (reasoning) { - if (thinking && !chunk.text.includes(EndThinkingToken)) { - thinking += chunk.text - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: chunk.text, - }) - } else { - // first time - if (!chunk.text.includes(StartThinkingToken)) { - let token = chunk.text - if (chunk.text.includes(EndThinkingToken)) { - token = chunk.text.split(EndThinkingToken)[0] - thinking += token - } else { - thinking += token - } - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: token, - }) - } - } - } - if (reasoning && chunk.text.includes(EndThinkingToken)) { - reasoning = false - chunk.text = chunk.text.split(EndThinkingToken)[1].trim() - } - if (!reasoning) { - answer += chunk.text - stream.writeSSE({ - event: ChatSSEvents.ResponseUpdate, - data: chunk.text, - }) - } - } + if (chunk.text) { + if (chunk.reasoning) { + thinking += chunk.text + stream.writeSSE({ + event: ChatSSEvents.Reasoning, + data: chunk.text, + }) + } else { + answer += chunk.text + stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: chunk.text, + }) + } + }
🧹 Nitpick comments (1)
server/api/chat/chat.ts (1)
223-224: Unify GroundingSupport typing to avoid driftPrefer importing GroundingSupport from a single internal type source to prevent version/shape drift with the provider SDK.
-import type { GroundingSupport } from "@google/genai" +import type { GroundingSupport } from "@/ai/types"
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (1)
server/api/chat/chat.ts(8 hunks)
🧰 Additional context used
🧬 Code graph analysis (1)
server/api/chat/chat.ts (3)
server/ai/types.ts (2)
WebSearchSource(95-99)GroundingSupport(101-108)server/shared/types.ts (3)
Citation(63-63)Apps(31-31)WebSearchEntity(36-36)server/ai/provider/index.ts (3)
webSearchQuestion(1793-1834)generateSearchQueryOrAnswerFromConversation(1457-1495)jsonParseLLMOutput(501-628)
🔇 Additional comments (1)
server/api/chat/chat.ts (1)
3866-3940: No duplicate implementations found—no shared util extraction needed
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (3)
server/package.json (3)
32-32: Optional: pin or vendor‑lock the SDK version to reduce surprise upgrades.Caret ranges on cloud SDKs can introduce breaking behavior via transitive updates. If stability > velocity, pin the exact version.
Apply within this line:
- "@google-cloud/vertexai": "^1.10.0", + "@google-cloud/vertexai": "1.10.0",
32-32: Operational follow‑ups: IAM, region, and timeouts.
- Ensure least-privileged SA and region scoping (e.g.,
GOOGLE_PROJECT_ID,GOOGLE_LOCATIONlikeus-central1).- Validate request timeouts/retries on Vertex calls and surface them via config.
- Add required vars to
.env.exampleand runtime validation.If not already present in the codebase, I can open a follow-up to add config schema checks for these envs.
32-32: Consolidate Google LLM SDK usage or clearly gate providers. We currently import both@google-cloud/vertexai(inserver/ai/provider/vertex_ai.ts) and@google/genai(inserver/ai/provider/index.tsandserver/api/chat/chat.ts), plus the Anthropic Vertex plugin. If you only need Vertex-backed Gemini, remove the@google/genaidependency and its GeminiAIProvider; otherwise, ensure each provider is cleanly gated and documented.
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (1)
server/package.json(1 hunks)
🔇 Additional comments (1)
server/package.json (1)
32-32: Add missing engines field & verify Bun compatibility for @google-cloud/vertexai
- The
server/package.jsonhas no"engines"block; add:"engines": { "node": ">=18", "bun": ">=1.0.30" }- Manually confirm that the ESM-only
@google-cloud/vertexaipackage (and its dependencies like google-auth-library/gaxios) runs correctly under Bun with no Node-specific gaps.
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (2)
server/search/vespa.ts (2)
867-869: Escape IDs in YQL to prevent injection and parsing issuesWrap interpolated IDs with
escapeYqlValue(...)(already defined in this file). This is both safer and more robust if an ID ever contains a quote.Apply this diff:
- const collectionCondition = `(${collectionIds.map((id: string) => `clId contains '${id.trim()}'`).join(" or ")})` + const collectionCondition = `(${collectionIds.map((id: string) => `clId contains '${escapeYqlValue(id.trim())}'`).join(" or ")})`- const folderCondition = `(${clVespaIds.map((id: string) => `docId contains '${id.trim()}'`).join(" or ")})` + const folderCondition = `(${clVespaIds.map((id: string) => `docId contains '${escapeYqlValue(id.trim())}'`).join(" or ")})`- const fileCondition = `(${clVespaIds.map((id: string) => `docId contains '${id.trim()}'`).join(" or ")})` + const fileCondition = `(${clVespaIds.map((id: string) => `docId contains '${escapeYqlValue(id.trim())}'`).join(" or ")})`Also applies to: 881-883, 895-897
904-911: Enforce ACL on KnowledgeBase Vespa query
In buildCollectionFileYQL’s returned YQL (around lines 904–911), append a permissions/ownership guard:and ${finalCondition} + and (permissions contains @email or owner contains @email)to restrict KB results to items the current user can access.
♻️ Duplicate comments (4)
server/search/vespa.ts (1)
900-901: Fix: fallback must not broaden scope; default to "false"If no collection/folder/file IDs resolve,
finalConditionshould be"false"to avoid unintentionally matching all KB items.Apply this diff:
- const finalCondition = - conditions.length > 0 ? `(${conditions.join(" or ")})` : "true" + const finalCondition = + conditions.length > 0 ? `(${conditions.join(" or ")})` : "false"server/ai/provider/index.ts (1)
1797-1843: webSearchQuestion: disable reasoning for Vertex GOOGLE and avoid per-call provider creation
- Vertex GOOGLE streaming doesn’t emit Start/End thinking markers; leaving
reasoningon misroutes output downstream. Forceparams.reasoning = falsehere.- A new
VertexAiProvideris created per call; initialize once and reuse (singleton), similar to other providers. This also avoids repeatedly constructing clients.+// at module scope with other provider singletons +let vertexGoogleProviderSingleton: VertexAiProvider | null = null// inside initializeProviders() after existing Vertex provider init if (VertexProjectId && VertexRegion) { // ... existing provider init ... + // Always prepare a GOOGLE-backed provider for web search + vertexGoogleProviderSingleton = new VertexAiProvider({ + projectId: VertexProjectId, + region: VertexRegion, + provider: VertexProvider.GOOGLE, + }) + Logger.info("Initialized VertexAI GOOGLE web-search provider singleton") }export const webSearchQuestion = ( query: string, userCtx: string, params: ModelParams, ): AsyncIterableIterator<ConverseResponse> => { try { if (!params.modelId) { params.modelId = defaultBestModel } - params.webSearch = true + params.webSearch = true + // Vertex GOOGLE doesn't emit thinking markers; avoid downstream gating. + params.reasoning = false if (!isAgentPromptEmpty(params.agentPrompt)) { params.systemPrompt = agentSearchQueryPrompt( userCtx, parseAgentPrompt(params.agentPrompt), ) } else if (!params.systemPrompt) { params.systemPrompt = "You are a helpful AI assistant with access to web search. Use web search when you need current information or real-time data to answer the user's question accurately." } @@ - if (!config.VertexProjectId || !config.VertexRegion) { + if (!config.VertexProjectId || !config.VertexRegion) { Logger.warn( "VertexProjectId/VertexRegion not configured, moving with default provider.", ) return getProviderByModel(params.modelId).converseStream(messages, params) } - const vertexGoogleProvider = new VertexAiProvider({ - projectId: config.VertexProjectId!, - region: config.VertexRegion!, - provider: VertexProvider.GOOGLE, - }) - - return vertexGoogleProvider.converseStream(messages, params) + if (!vertexGoogleProviderSingleton) { + throw new Error("Vertex GOOGLE provider not initialized") + } + return vertexGoogleProviderSingleton.converseStream(messages, params)server/api/chat/chat.ts (2)
3908-3982: Fix citation insertion: sort by endIndex descending to avoid shifting; map using citationIndexInserting “[n]” left-to-right changes subsequent indices. Insert right-to-left or track offsets. Also record
newCitationMapwithcitationIndex(notsourceIndex) to avoid mismaps when reusing indices.function processWebSearchCitations( answer: string, allSources: WebSearchSource[], finalGroundingSupports: GroundingSupport[], citations: Citation[], citationMap: Record<number, number>, sourceIndex: number, ): { @@ - if (finalGroundingSupports.length > 0 && allSources.length > 0) { + if (finalGroundingSupports.length > 0 && allSources.length > 0) { let answerWithCitations = answer let newCitations: Citation[] = [] let newCitationMap: Record<number, number> = {} let urlToIndexMap: Map<string, number> = new Map() - for (const support of finalGroundingSupports) { + // Insert from right to left to avoid index shifting + const supportsSorted = [...finalGroundingSupports].sort( + (a, b) => (b.segment?.endIndex ?? 0) - (a.segment?.endIndex ?? 0), + ) + for (const support of supportsSorted) { const segment = support.segment const groundingChunkIndices = support.groundingChunkIndices || [] let citationText = "" for (const chunkIndex of groundingChunkIndices) { if (allSources[chunkIndex]) { const source = allSources[chunkIndex] let citationIndex: number if (urlToIndexMap.has(source.uri)) { // Reuse existing citation index citationIndex = urlToIndexMap.get(source.uri)! } else { citationIndex = sourceIndex const webSearchCitation: Citation = { docId: `websearch_${sourceIndex}`, title: source.title, url: source.uri, app: Apps.WebSearch, entity: WebSearchEntity.WebSearch, } newCitations.push(webSearchCitation) - newCitationMap[sourceIndex] = - citations.length + newCitations.length - 1 + newCitationMap[citationIndex] = + citations.length + newCitations.length - 1 urlToIndexMap.set(source.uri, citationIndex) sourceIndex++ } citationText += ` [${citationIndex}]` } } if ( citationText && segment?.endIndex !== undefined && segment.endIndex <= answerWithCitations.length ) { answerWithCitations = answerWithCitations.slice(0, segment.endIndex) + citationText + answerWithCitations.slice(segment.endIndex) } } return { updatedAnswer: answerWithCitations, newCitations, newCitationMap, updatedSourceIndex: sourceIndex, } } return null }Also consider de-duping this helper (appears twice per PR summary) into a shared util to avoid divergence.
4587-4630: Web-search streaming: route by chunk.reasoning and pass reasoning=false to Vertex
- Current branch gates on Start/EndThinking tokens; Vertex GOOGLE doesn’t emit them, so output may be swallowed/misrouted.
- Use
chunk.reasoningwhen present; otherwise treat all text as response. Also callwebSearchQuestionwithreasoning: false.- searchOrAnswerIterator = webSearchQuestion(message, ctx, { + searchOrAnswerIterator = webSearchQuestion(message, ctx, { modelId: Models.Gemini_2_5_Flash, stream: true, - json: false, + json: false, agentPrompt: agentPromptValue, - reasoning: - userRequestsReasoning && - ragPipelineConfig[RagPipelineStages.AnswerOrSearch].reasoning, + // Vertex GOOGLE: disable local reasoning gating + reasoning: false, messages: llmFormattedMessages, webSearch: true, })- if (chunk.text) { - if (reasoning) { - if (thinking && !chunk.text.includes(EndThinkingToken)) { - thinking += chunk.text - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: chunk.text, - }) - } else { - // first time - if (!chunk.text.includes(StartThinkingToken)) { - let token = chunk.text - if (chunk.text.includes(EndThinkingToken)) { - token = chunk.text.split(EndThinkingToken)[0] - thinking += token - } else { - thinking += token - } - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: token, - }) - } - } - } - if (reasoning && chunk.text.includes(EndThinkingToken)) { - reasoning = false - chunk.text = chunk.text.split(EndThinkingToken)[1].trim() - } - if (!reasoning) { - answer += chunk.text - stream.writeSSE({ - event: ChatSSEvents.ResponseUpdate, - data: chunk.text, - }) - } - } + if (chunk.text) { + if (chunk.reasoning) { + thinking += chunk.text + stream.writeSSE({ + event: ChatSSEvents.Reasoning, + data: chunk.text, + }) + } else { + answer += chunk.text + stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: chunk.text, + }) + } + }Also applies to: 4669-4789
🧹 Nitpick comments (4)
server/search/vespa.ts (1)
852-862: Deduplicate and sanitize input IDs before building conditionsAvoid redundant OR terms and trim early; this keeps YQL smaller and faster.
Apply this diff:
for (const selection of collectionSelections) { if (selection.collectionIds) { collectionIds.push(...selection.collectionIds) } if (selection.collectionFolderIds) { collectionFolderIds.push(...selection.collectionFolderIds) } if (selection.collectionFileIds) { collectionFileIds.push(...selection.collectionFileIds) } } + // Deduplicate + trim + const uniqueCollectionIds = [...new Set(collectionIds.map((v) => v?.trim()).filter(Boolean))] + const uniqueCollectionFolderIds = [...new Set(collectionFolderIds.map((v) => v?.trim()).filter(Boolean))] + const uniqueCollectionFileIds = [...new Set(collectionFileIds.map((v) => v?.trim()).filter(Boolean))]Then update subsequent checks to use the deduped arrays:
- if (collectionIds.length > 0) { - const collectionCondition = `(${collectionIds.map((id: string) => `clId contains '${escapeYqlValue(id.trim())}'`).join(" or ")})` + if (uniqueCollectionIds.length > 0) { + const collectionCondition = `(${uniqueCollectionIds.map((id: string) => `clId contains '${escapeYqlValue(id)}'`).join(" or ")})`- if (collectionFolderIds.length > 0) { - const clFileIds = await getAllFolderItems(collectionFolderIds, db) + if (uniqueCollectionFolderIds.length > 0) { + const clFileIds = await getAllFolderItems(uniqueCollectionFolderIds, db)- if (collectionFileIds.length > 0) { - const ids = await getCollectionFilesVespaIds(collectionFileIds, db) + if (uniqueCollectionFileIds.length > 0) { + const ids = await getCollectionFilesVespaIds(uniqueCollectionFileIds, db)server/ai/provider/index.ts (1)
282-290: VERTEX_PROVIDER mapping: log enum name, not numeric; handle invalid values explicitly
- Current log will print a number (enum value). Prefer the enum key for readability.
- Optional: log a warning when
VERTEX_PROVIDERis set but invalid (you already fall back).- const provider = + const provider = vertexProviderType && VertexProvider[vertexProviderType] ? VertexProvider[vertexProviderType] : VertexProvider.ANTHROPIC - Logger.info(`Initialized VertexAI provider with ${provider} backend`) + const providerName = + vertexProviderType && VertexProvider[vertexProviderType] + ? vertexProviderType + : "ANTHROPIC" + Logger.info(`Initialized VertexAI provider with ${providerName} backend`)Also applies to: 297-297
server/api/chat/chat.ts (2)
4547-4568: Parsing prior classification: optional robustnessMinor: narrow the try/catch to just JSON.parse and guard with a cheap startsWith("{") before parsing to reduce noise in logs. Current code is otherwise fine.
5531-5531: Variable shadowing nit
const prevUserMessageis declared inside thisifblock and again later in the outer scope. Rename the inner one (e.g.,prevUserMsgForAttachments) to avoid confusion.- const prevUserMessage = conversation[conversation.length - 1] + const prevUserMsgForAttachments = conversation[conversation.length - 1] if (prevUserMessage.messageRole === "user") { - attachmentMetadata = await getAttachmentsByMessageId( + attachmentMetadata = await getAttachmentsByMessageId( db, - prevUserMessage.externalId, + prevUserMsgForAttachments.externalId, email, )
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (5)
server/ai/provider/index.ts(3 hunks)server/ai/types.ts(1 hunks)server/api/chat/chat.ts(14 hunks)server/package.json(1 hunks)server/search/vespa.ts(2 hunks)
🚧 Files skipped from review as they are similar to previous changes (2)
- server/ai/types.ts
- server/package.json
🧰 Additional context used
🧬 Code graph analysis (2)
server/ai/provider/index.ts (3)
server/ai/provider/vertex_ai.ts (1)
VertexAiProvider(67-485)server/ai/types.ts (2)
ModelParams(69-84)ConverseResponse(86-93)server/ai/agentPrompts.ts (1)
agentSearchQueryPrompt(1026-1346)
server/api/chat/chat.ts (4)
server/ai/types.ts (3)
WebSearchSource(95-99)GroundingSupport(101-108)QueryRouterLLMResponse(212-212)server/api/chat/utils.ts (2)
getRecentChainBreakClassifications(917-922)formatChainBreaksForPrompt(958-973)server/db/schema/messages.ts (2)
messages(31-75)selectMessageSchema(84-84)server/ai/provider/index.ts (3)
webSearchQuestion(1797-1843)generateSearchQueryOrAnswerFromConversation(1459-1499)jsonParseLLMOutput(503-630)
🔇 Additional comments (6)
server/ai/provider/index.ts (1)
98-98: Expose VertexProvider: LGTMRe-exporting
VertexProvideris fine and unblocks callers.server/api/chat/chat.ts (5)
22-22: Type-safe web search wiring: LGTMImports for
webSearchQuestion,WebSearchSource,WebSearchEntity, andGroundingSupportlook correct.Also applies to: 40-41, 143-144, 228-229
2909-2913: Pass-through to processIterator: LGTMThe added args keep semantics consistent with other callsites.
5936-5957: Retry path chain-break/prev-classification: LGTMGood reuse of prior classification and localized chain-break window.
Also applies to: 5963-5971
6147-6149: Classification filters defaults: LGTMUsing
parsed.filters.offset || 0andintent || {}is reasonable.
4013-4014: Ignore chatSchema suggestion enableWebSearch is already defined in messageSchema (and inferred by MessageReqType), so it will be parsed correctly.Likely an incorrect or invalid review comment.
There was a problem hiding this comment.
Actionable comments posted: 0
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
server/api/chat/chat.ts (1)
4081-4082: Block-scoped redeclaration: ‘chat’ is declared twice in MessageApiYou declare “let chat: SelectChat” at Line 3998 and again at Line 4082 in the same try block; TS will error.
Apply:
- let chat: SelectChat + // reuse the top-level 'chat' declared earlier in this scope
♻️ Duplicate comments (2)
server/api/chat/chat.ts (2)
3909-3988: Fix citation insertion index shifting (insert right-to-left and de-dup per segment)Inserting “[n]” left-to-right mutates subsequent indices; repeated URLs within a segment can also duplicate “[n]”. Sort by segment.endIndex descending and build a unique set per segment.
Apply:
- for (const support of finalGroundingSupports) { + const supportsSorted = [...finalGroundingSupports].sort( + (a, b) => (b.segment?.endIndex ?? 0) - (a.segment?.endIndex ?? 0), + ) + for (const support of supportsSorted) { const segment = support.segment const groundingChunkIndices = support.groundingChunkIndices || [] - let citationText = "" + const citationIndices: number[] = [] for (const chunkIndex of groundingChunkIndices) { if (allSources[chunkIndex]) { const source = allSources[chunkIndex] let citationIndex: number if (urlToIndexMap.has(source.uri)) { // Reuse existing citation index citationIndex = urlToIndexMap.get(source.uri)! } else { citationIndex = sourceIndex const webSearchCitation: Citation = { docId: `websearch_${sourceIndex}`, title: source.title, url: source.uri, app: Apps.WebSearch, entity: WebSearchEntity.WebSearch, } newCitations.push(webSearchCitation) - newCitationMap[sourceIndex] = - citations.length + newCitations.length - 1 + newCitationMap[citationIndex] = + citations.length + newCitations.length - 1 urlToIndexMap.set(source.uri, sourceIndex) sourceIndex++ } - citationText += ` [${citationIndex}]` + citationIndices.push(citationIndex) } } + const uniqueCitationIndices = [...new Set(citationIndices)] + const citationText = + uniqueCitationIndices.length > 0 + ? uniqueCitationIndices.map((i) => ` [${i}]`).join("") + : "" if ( citationText && segment?.endIndex !== undefined && segment.endIndex <= answerWithCitations.length ) { // Find optimal insertion point that respects word boundaries const optimalIndex = findOptimalCitationInsertionPoint( answerWithCitations, segment.endIndex, ) answerWithCitations = answerWithCitations.slice(0, optimalIndex) + citationText + answerWithCitations.slice(optimalIndex) } }#!/bin/bash # Ensure only a single implementation exists or update all copies. rg -nP 'function\s+processWebSearchCitations\s*\(' -C2
4699-4734: Web-search streaming: all tokens routed to Reasoning (Vertex doesn’t emit Start/EndThinkingToken)This branch gates on Start/EndThinkingToken; Vertex chunks typically set chunk.reasoning instead. Today, reasoning may never flip to false → no ResponseUpdate.
Apply:
- if (chunk.text) { - if (reasoning) { - if (thinking && !chunk.text.includes(EndThinkingToken)) { - thinking += chunk.text - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: chunk.text, - }) - } else { - // first time - if (!chunk.text.includes(StartThinkingToken)) { - let token = chunk.text - if (chunk.text.includes(EndThinkingToken)) { - token = chunk.text.split(EndThinkingToken)[0] - thinking += token - } else { - thinking += token - } - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: token, - }) - } - } - } - if (reasoning && chunk.text.includes(EndThinkingToken)) { - reasoning = false - chunk.text = chunk.text.split(EndThinkingToken)[1].trim() - } - if (!reasoning) { - answer += chunk.text - stream.writeSSE({ - event: ChatSSEvents.ResponseUpdate, - data: chunk.text, - }) - } - } + if (chunk.text) { + if (chunk.reasoning) { + thinking += chunk.text + stream.writeSSE({ + event: ChatSSEvents.Reasoning, + data: chunk.text, + }) + } else { + answer += chunk.text + stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: chunk.text, + }) + } + }
🧹 Nitpick comments (1)
server/api/chat/utils.ts (1)
1001-1045: Comment contradicts implementation; clarify or adjust tie-breakCode favors the left boundary on ties, but the comment says “lean towards right.” Either update the comment or flip the conditional.
Apply one of:
- // Prefer the closer boundary, but lean towards right boundary (end of word) for better readability - if (leftDistance <= rightDistance || rightBoundary >= text.length) { + // Prefer the closer boundary; lean towards left boundary on ties + if (leftDistance <= rightDistance || rightBoundary >= text.length) { return leftBoundary } else { return rightBoundary }or, to actually lean right:
- if (leftDistance <= rightDistance || rightBoundary >= text.length) { + if (rightDistance <= leftDistance || rightBoundary >= text.length) { return leftBoundary } else { return rightBoundary }
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (2)
server/api/chat/chat.ts(15 hunks)server/api/chat/utils.ts(3 hunks)
🧰 Additional context used
🧬 Code graph analysis (2)
server/api/chat/chat.ts (4)
server/ai/types.ts (2)
WebSearchSource(95-99)GroundingSupport(101-108)server/api/chat/types.ts (1)
Citation(76-76)server/api/chat/utils.ts (1)
findOptimalCitationInsertionPoint(1001-1045)server/ai/provider/index.ts (1)
webSearchQuestion(1797-1843)
server/api/chat/utils.ts (3)
server/ai/types.ts (1)
QueryRouterLLMResponse(212-212)server/db/schema/messages.ts (2)
messages(31-75)SelectMessage(85-85)server/logger/index.ts (1)
getLoggerWithChild(189-197)
🔇 Additional comments (8)
server/api/chat/utils.ts (5)
888-892: ChainBreakClassification: style-only change looks goodNo behavioral impact; types remain the same.
904-908: Stronger type guard—goodRejecting non-object/array payloads prevents misclassification noise.
921-930: Recent chain-break limiter—goodTaking the two most recent entries is sensible for prompt budget.
933-979: Chain-break extraction logic—OK with implicit “user/assistant/user” strideAssumes alternating roles (index - 2). If mixed-role sequences are possible (e.g., tool messages), consider walking back to the previous “user” instead of index math.
981-999: Null vs object return—goodReturning null when none exist keeps prompts lean.
server/api/chat/chat.ts (3)
2910-2914: Good: unified streaming path for metadata answersPassing items and reasoning flag through processIterator keeps behavior consistent.
4593-4636: Feature flag plumbing—OKwebSearchEnabled path is correctly gated and preserves the non-web path.
6153-6155: Retry classification: safe defaultsUsing parsed.filters.offset || 0 and intent || {} avoids undefineds.
There was a problem hiding this comment.
Actionable comments posted: 2
♻️ Duplicate comments (3)
server/api/chat/chat.ts (3)
229-231: Nice: typed GroundingSupport/WebSearchSource improve safetyImports and function signature now use precise types, removing any[]. LGTM.
Also applies to: 3911-3915
3909-3988: Citations can be inserted at wrong places due to shifting indices; insert right-to-leftInserting “[n]” mutates answer length and invalidates subsequent segment.endIndex. Sort supports by endIndex desc (or track cumulative offset). Also map using citationIndex for clarity.
function processWebSearchCitations( answer: string, allSources: WebSearchSource[], finalGroundingSupports: GroundingSupport[], citations: Citation[], citationMap: Record<number, number>, sourceIndex: number, ): { @@ - if (finalGroundingSupports.length > 0 && allSources.length > 0) { + if (finalGroundingSupports.length > 0 && allSources.length > 0) { let answerWithCitations = answer let newCitations: Citation[] = [] let newCitationMap: Record<number, number> = {} let urlToIndexMap: Map<string, number> = new Map() - for (const support of finalGroundingSupports) { + // Insert from right to left to avoid index shifting + const supportsSorted = [...finalGroundingSupports].sort( + (a, b) => (b.segment?.endIndex ?? 0) - (a.segment?.endIndex ?? 0), + ) + for (const support of supportsSorted) { const segment = support.segment const groundingChunkIndices = support.groundingChunkIndices || [] let citationText = "" for (const chunkIndex of groundingChunkIndices) { if (allSources[chunkIndex]) { const source = allSources[chunkIndex] let citationIndex: number if (urlToIndexMap.has(source.uri)) { // Reuse existing citation index citationIndex = urlToIndexMap.get(source.uri)! } else { citationIndex = sourceIndex const webSearchCitation: Citation = { docId: `websearch_${sourceIndex}`, title: source.title, url: source.uri, app: Apps.WebSearch, entity: WebSearchEntity.WebSearch, } newCitations.push(webSearchCitation) - newCitationMap[sourceIndex] = - citations.length + newCitations.length - 1 - urlToIndexMap.set(source.uri, sourceIndex) + newCitationMap[citationIndex] = + citations.length + newCitations.length - 1 + urlToIndexMap.set(source.uri, citationIndex) sourceIndex++ } citationText += ` [${citationIndex}]` } } if ( citationText && segment?.endIndex !== undefined && segment.endIndex <= answerWithCitations.length ) { // Find optimal insertion point that respects word boundaries const optimalIndex = findOptimalCitationInsertionPoint( answerWithCitations, segment.endIndex, ) answerWithCitations = answerWithCitations.slice(0, optimalIndex) + citationText + answerWithCitations.slice(optimalIndex) } }
4698-4734: Vertex web-search streaming: all tokens go to Reasoning; no ResponseUpdate emittedThis branch relies on Start/EndThinkingToken, which Vertex web search doesn’t emit. Result: ResponseUpdate is starved.
- if (chunk.text) { - if (reasoning) { - if (thinking && !chunk.text.includes(EndThinkingToken)) { - thinking += chunk.text - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: chunk.text, - }) - } else { - // first time - if (!chunk.text.includes(StartThinkingToken)) { - let token = chunk.text - if (chunk.text.includes(EndThinkingToken)) { - token = chunk.text.split(EndThinkingToken)[0] - thinking += token - } else { - thinking += token - } - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: token, - }) - } - } - } - if (reasoning && chunk.text.includes(EndThinkingToken)) { - reasoning = false - chunk.text = chunk.text.split(EndThinkingToken)[1].trim() - } - if (!reasoning) { - answer += chunk.text - stream.writeSSE({ - event: ChatSSEvents.ResponseUpdate, - data: chunk.text, - }) - } - } + if (chunk.text) { + if (chunk.reasoning) { + thinking += chunk.text + await stream.writeSSE({ + event: ChatSSEvents.Reasoning, + data: chunk.text, + }) + } else { + answer += chunk.text + await stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: chunk.text, + }) + } + }
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (2)
server/ai/provider/index.ts(3 hunks)server/api/chat/chat.ts(16 hunks)
🚧 Files skipped from review as they are similar to previous changes (1)
- server/ai/provider/index.ts
🧰 Additional context used
🧬 Code graph analysis (1)
server/api/chat/chat.ts (4)
server/ai/types.ts (3)
WebSearchSource(95-99)GroundingSupport(101-108)QueryRouterLLMResponse(212-212)server/api/chat/types.ts (1)
Citation(76-76)server/api/chat/utils.ts (1)
findOptimalCitationInsertionPoint(1001-1045)server/ai/provider/index.ts (2)
webSearchQuestion(1797-1841)generateSearchQueryOrAnswerFromConversation(1459-1499)
There was a problem hiding this comment.
Actionable comments posted: 1
♻️ Duplicate comments (4)
server/api/chat/chat.ts (4)
4019-4023: Plumbing: ensure enableWebSearch is in the validated request schema and frontendThe handler reads enableWebSearch from the request; verify the schema (message-level) and types include it, and that the client sends it.
#!/bin/bash # Verify schema/type plumbing and call sites for enableWebSearch rg -nP 'messageSchema|MessageReqType' server/api/search.ts -C3 || true rg -n 'enableWebSearch' -- server frontend || trueAlso applies to: 4593-4636
3909-3988: Fix index drift: insert citations from right-to-leftEach insertion shifts subsequent indices; sort by segment.endIndex descending before splicing. Also map using citationIndex for clarity.
- for (const support of finalGroundingSupports) { + const supportsSorted = [...finalGroundingSupports].sort( + (a, b) => (b.segment?.endIndex ?? 0) - (a.segment?.endIndex ?? 0), + ) + for (const support of supportsSorted) { const segment = support.segment const groundingChunkIndices = support.groundingChunkIndices || [] let citationText = "" for (const chunkIndex of groundingChunkIndices) { if (allSources[chunkIndex]) { const source = allSources[chunkIndex] let citationIndex: number if (urlToIndexMap.has(source.uri)) { // Reuse existing citation index citationIndex = urlToIndexMap.get(source.uri)! } else { citationIndex = sourceIndex const webSearchCitation: Citation = { docId: `websearch_${sourceIndex}`, title: source.title, url: source.uri, app: Apps.WebSearch, entity: WebSearchEntity.WebSearch, } newCitations.push(webSearchCitation) - newCitationMap[sourceIndex] = - citations.length + newCitations.length - 1 + newCitationMap[citationIndex] = + citations.length + newCitations.length - 1 urlToIndexMap.set(source.uri, sourceIndex) sourceIndex++ } citationText += ` [${citationIndex}]` } }
4698-4734: Streaming reasoning gating for web search should use chunk.reasoningVertex (GOOGLE) doesn’t emit Start/EndThinkingToken. Route tokens via chunk.reasoning so answers don’t get stuck in “Reasoning.”
- if (chunk.text) { - if (reasoning) { - if (thinking && !chunk.text.includes(EndThinkingToken)) { - thinking += chunk.text - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: chunk.text, - }) - } else { - // first time - if (!chunk.text.includes(StartThinkingToken)) { - let token = chunk.text - if (chunk.text.includes(EndThinkingToken)) { - token = chunk.text.split(EndThinkingToken)[0] - thinking += token - } else { - thinking += token - } - stream.writeSSE({ - event: ChatSSEvents.Reasoning, - data: token, - }) - } - } - } - if (reasoning && chunk.text.includes(EndThinkingToken)) { - reasoning = false - chunk.text = chunk.text.split(EndThinkingToken)[1].trim() - } - if (!reasoning) { - answer += chunk.text - stream.writeSSE({ - event: ChatSSEvents.ResponseUpdate, - data: chunk.text, - }) - } - } + if (chunk.text) { + if (chunk.reasoning) { + thinking += chunk.text + stream.writeSSE({ + event: ChatSSEvents.Reasoning, + data: chunk.text, + }) + } else { + answer += chunk.text + stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: chunk.text, + }) + } + }
2910-2914: Bug: Missing email in processIterator causes image-citation failuresprocessIterator dereferences email when emitting image citations; passing undefined can break streaming.
return yield* processIterator( iterator, items, 0, config.isReasoning && userRequestsReasoning, + email, )
🧹 Nitpick comments (1)
server/api/chat/chat.ts (1)
229-229: Use internal GroundingSupport type to decouple from provider SDKPrefer the shared interface from "@/ai/types" to avoid lock-in to @google/genai shapes and to keep server types consistent.
- import type { GroundingSupport } from "@google/genai" + import type { GroundingSupport } from "@/ai/types"
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (2)
server/api/chat/chat.ts(16 hunks)server/api/search.ts(1 hunks)
🚧 Files skipped from review as they are similar to previous changes (1)
- server/api/search.ts
🧰 Additional context used
🧬 Code graph analysis (1)
server/api/chat/chat.ts (4)
server/ai/types.ts (2)
WebSearchSource(95-99)GroundingSupport(101-108)server/api/chat/types.ts (1)
Citation(76-76)server/api/chat/utils.ts (1)
findOptimalCitationInsertionPoint(1001-1045)server/ai/provider/index.ts (1)
webSearchQuestion(1797-1841)
🔇 Additional comments (1)
server/api/chat/chat.ts (1)
205-206: Good call: word-boundary-safe citation insertion helperImporting and using findOptimalCitationInsertionPoint improves readability and avoids mid-word insertions.
There was a problem hiding this comment.
Actionable comments posted: 0
♻️ Duplicate comments (3)
server/api/chat/chat.ts (3)
2910-2914: Pass email into processIterator to prevent image-citation streaming bugsprocessIterator dereferences email when emitting image citations. Omit -> runtime errors or missing images.
Apply:
return yield* processIterator( iterator, items, 0, - config.isReasoning && userRequestsReasoning, + config.isReasoning && userRequestsReasoning, + email, )
4706-4723: Grounding index/source mismatch: align sources array with final supportsYou dedupe into allSources across chunks, but the final groundingChunkIndices refer to the final chunk’s sources array. This can point to the wrong URL/title.
- if (chunk.sources && chunk.sources.length > 0) { + if (chunk.sources && chunk.sources.length > 0) { chunk.sources.forEach((source) => { if ( !allSources.some( (existing) => existing.uri === source.uri, ) ) { allSources.push(source) } }) } if ( chunk.groundingSupports && chunk.groundingSupports.length > 0 ) { - finalGroundingSupports = chunk.groundingSupports + finalGroundingSupports = chunk.groundingSupports + // IMPORTANT: keep indices aligned with the provider's final sources + if (chunk.sources && chunk.sources.length > 0) { + allSources = chunk.sources + } }
3909-3988: Citation insertion order can corrupt indices; insert right-to-leftEach insertion shifts subsequent endIndex positions. Sort supports by segment.endIndex desc (or track cumulative offset).
Apply:
- for (const support of finalGroundingSupports) { + // Insert from right to left to avoid index shifting + const supportsSorted = [...finalGroundingSupports].sort( + (a, b) => (b.segment?.endIndex ?? 0) - (a.segment?.endIndex ?? 0), + ) + for (const support of supportsSorted) { const segment = support.segment const groundingChunkIndices = support.groundingChunkIndices || [] let citationText = "" for (const chunkIndex of groundingChunkIndices) { if (allSources[chunkIndex]) { const source = allSources[chunkIndex] let citationIndex: number if (urlToIndexMap.has(source.uri)) { // Reuse existing citation index citationIndex = urlToIndexMap.get(source.uri)! } else { citationIndex = sourceIndex const webSearchCitation: Citation = { docId: `websearch_${sourceIndex}`, title: source.title, url: source.uri, app: Apps.WebSearch, entity: WebSearchEntity.WebSearch, } newCitations.push(webSearchCitation) - newCitationMap[sourceIndex] = - citations.length + newCitations.length - 1 + newCitationMap[citationIndex] = + citations.length + newCitations.length - 1 urlToIndexMap.set(source.uri, sourceIndex) sourceIndex++ } citationText += ` [${citationIndex}]` } } if ( citationText && segment?.endIndex !== undefined && segment.endIndex <= answerWithCitations.length ) { // Find optimal insertion point that respects word boundaries const optimalIndex = findOptimalCitationInsertionPoint( answerWithCitations, segment.endIndex, ) answerWithCitations = answerWithCitations.slice(0, optimalIndex) + citationText + answerWithCitations.slice(optimalIndex) } }
🧹 Nitpick comments (2)
server/api/chat/chat.ts (2)
4697-4705: Stream reasoning tokens if provider sets chunk.reasoningVertex GOOGLE may not set Start/EndThinkingToken. Prefer per-chunk flag when present; today all text goes to ResponseUpdate.
- if (chunk.text) { - answer += chunk.text - stream.writeSSE({ - event: ChatSSEvents.ResponseUpdate, - data: chunk.text, - }) - } + if (chunk.text) { + if (chunk.reasoning) { + await stream.writeSSE({ + event: ChatSSEvents.Reasoning, + data: chunk.text, + }) + } else { + answer += chunk.text + await stream.writeSSE({ + event: ChatSSEvents.ResponseUpdate, + data: chunk.text, + }) + } + }
5507-5517: Shadowed variable name reduces clarityprevUserMessage here shadows the outer prevUserMessage defined later in the function. Rename to avoid confusion.
- const prevUserMessage = conversation[conversation.length - 1] + const lastConvUserMessage = conversation[conversation.length - 1]
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
Plan: Pro
💡 Knowledge Base configuration:
- MCP integration is disabled by default for public repositories
- Jira integration is disabled by default for public repositories
- Linear integration is disabled by default for public repositories
You can enable these sources in your CodeRabbit configuration.
📒 Files selected for processing (1)
server/api/chat/chat.ts(16 hunks)
🧰 Additional context used
🧬 Code graph analysis (1)
server/api/chat/chat.ts (4)
server/ai/types.ts (3)
WebSearchSource(95-99)GroundingSupport(101-108)QueryRouterLLMResponse(212-212)server/api/chat/types.ts (2)
Citation(76-76)ragPipelineConfig(35-63)server/api/chat/utils.ts (1)
findOptimalCitationInsertionPoint(1001-1045)server/ai/provider/index.ts (2)
webSearchQuestion(1797-1841)generateSearchQueryOrAnswerFromConversation(1459-1499)
🔇 Additional comments (2)
server/api/chat/chat.ts (2)
229-229: Use local GroundingSupport in chat.ts
Line 229 currently imports GroundingSupport from “@google/genai” but the project defines its own interface at server/ai/types.ts:101. Confirm the external and local definitions match exactly—or replace the external import with the local type to avoid runtime typing mismatches.
4593-4611: enableWebSearch is plumbed correctly. The server’smessageSchemaincludes anenableWebSearchstring-to-boolean transform andMessageReqTypeexposes it, and the frontend always appendsenableWebSearchto the request URL.
3 participants
Description
Testing
Additional Notes
Summary by CodeRabbit
New Features
Improvements
Refactor
Chores