The purpose of this space is to explain and demostrate the potential of AI Agents using entirely Typescript
An AI Agent is a piece of software that uses AI for executing tasks.
The key difference between agent versus other software is the ability of reasoning and acting based on external information. An example of this behaviour is the use of ReAct framework, which is described in ReAct Framework.
Some of the components of an AI Agent are: planning, memory, tools and action.
Anthropic defines two main categories:
- Workflow: Systems where the flow is controlled by a pre-defined path.
- Agents: Systems where LLMs defines the flow dynamically.
- Prompt chainning: The final result is obtained by splitting a major work into smaller tasks, which are executed by LLMs in sequence.
- Routing: The LLM decides the next step to execute in a flow.
- Parallelization: A software like python, js, etc orchestrates multiple LLMs to execute task in parallel to solve a problem.
- Orchestrator-worker: A LLM defines the tasks and orquestate multiple LLMs to execute the tasks in parallel to solve a problem.
- Evaluator-Optimizer: A LLM evaluates the output of other LLMs and provides feedback to improve the result.
A tool is a function schema, which contains information such as name; description; parameters, and It is available to the LLM. Tool calling is the ability of a LLM to invoke the tools in order to interact with the environment, for example: get current time, fetch meetings in a calendar, etc.
This process decompone the work into smaller steps & actions. There are several strategies for planning, such as:
- Chain-of-Thought (CoT): Enhances the reasoning capability by decomposing the task in several steps, which are executed in sequence in order to elaborate an output.
- Tree of Through reasoning (ToT): This strategy split the work in several paths and tasks. Unlike CoT strategy, the tasks are executed in parallel, since multiple paths are evaluated at the same time.
- Planning with Feedback / ReAct (Reasoning and Act): Described in ReAct Framework.
This component allows AI Agents to retain information about past experience in order to learn, adapt and make better decisions.
External funcionalities, such as REST APIs, that let the AI Agent interact with external entities.
The task that is executed by the Agent, for example reserving a meeting by using a tool.
The Resoning and Acting (ReAct) framework describes a technique for leveraging LLMs models by using reasoning and acting strategies. The approach consist of prompt a LLM in order to generate a list of "thoughts", where a though is a reasoning, in order to execute tasks (acting) that might be execute APIs, query to databases, etc.
| Name | Provider | Description | Best For | Pros | Cons |
|---|---|---|---|---|---|
| Vercel AI SDK | Vercel | A TypeScript toolkit for building AI applications and agents with support for multiple frameworks like Next.js, React, and Svelte. | Building AI-powered web applications with seamless UI integration. | Unified API for multiple LLM providers, strong streaming capabilities, framework-agnostic UI hooks, robust community support. | Limited to TypeScript/JavaScript ecosystems, less focus on complex multi-agent workflows. |
| AutoGen | Microsoft | A framework for developing multi-agent LLM applications with conversational orchestration. | Multi-agent systems requiring dynamic conversations and task collaboration. | Supports diverse tools, strong Microsoft backing, flexible for conversational workflows, good for research and prototyping. | Complex interaction management, not fully production-ready, steeper learning curve for non-conversational use cases. |
| OpenAI Agents SDK | OpenAI | A structured toolset for building agents that reason, plan, and call external APIs, integrated with OpenAI's models. | Developers already using OpenAI's ecosystem for agent-based applications. | Simple orchestration, strong integration with OpenAI models, good for agent handoffs, Python-based for easy adoption. | Lacks built-in parallel execution, experimental features may change, limited to OpenAI stack. |
| LangChain | LangChain | A comprehensive framework for chaining LLMs with tools, memory, and data sources, with support for RAG and agent workflows. | Complex applications needing RAG, tool integration, or multi-step workflows. | Extensive integrations, large community, supports multiple LLMs and vector stores, flexible for custom workflows. | Can be overly abstract for simple use cases, relies on third-party data storage, steep learning curve for advanced features. |
| LlamaIndex | LlamaIndex | A data framework for building LLM applications with efficient data ingestion, indexing, and retrieval, enhanced with agent capabilities. | Data-heavy tasks like question answering or retrieval-centric applications. | Intuitive for data indexing, supports diverse data sources, seamless RAG integration, good for private document workflows. | Less focus on multi-agent orchestration, requires boilerplate for complex workflows, still evolving agent features. |
- Node.js and npm/yarn installed.
- Ollama installed and running. You can download it from ollama.com.
- At least one chat model (e.g.,
llama2) and one embedding model (e.g.,nomic-embed-text) pulled via Ollama:ollama pull llama2 ollama pull nomic-embed-text
- A Pinecone account and an API key. You will also need to have created a Pinecone index.
-
Clone the repository (if applicable).
-
Install dependencies:
npm install # or yarn install -
Configure Environment Variables: Create a
.envfile in the00_langchain_ollama_pyneconedirectory with the following content, replacing the placeholder values with your actual credentials and desired settings:PINECONE_API_KEY="YOUR_PINECONE_API_KEY" PINECONE_INDEX_NAME="your-pinecone-index-name" # Optional: Override default Ollama models and API URL OLLAMA_MODEL="llama2" OLLAMA_EMBEDDINGS_MODEL="nomic-embed-text:latest" # Model for generating embeddings OLLAMA_EMBEDDINGS_URL="http://localhost:11434/api/embeddings" # Ollama embeddings API endpoint
-
Prepare your TypeScript code: Place the TypeScript files you want to query into a directory. By default,
ingest.tslooks for a directory namedyour-typescript-code-directoryin the same location asingest.ts. You can change this path iningest.tsif needed.