Agent Flow

Agent Flow is a agent-based system built using the uagents framework. It combines document retrieval, SQL querying, and large language model reasoning to answer natural language questions. This system routes queries intelligently across different specialized agents to provide accurate, explainable, and context-aware responses.

Click to watch demo video for round 1

Click to watch demo video for round 2

Final presentation link

Overview

The system consists of several agents, each with a specific responsibility:

• Intent Classifier Agent: Determines the type of user query (SQL, document, hybrid, or unknown).
• SQL Agent: Translates natural language into SQL, executes it on a SQLite database (project.db), and summarizes the result.
• Document Agent: Performs similarity search across document embeddings to answer context-based questions.
• Hybrid Agent: Uses both similarity search and SQL execution for complex questions requiring multi-source reasoning.
• Error Handler Agent: Handles misclassified or ambiguous queries using LLM's general knowledge capabilities.

Agent Workflow Chart

Features

• Modular agent architecture using uagents.
• SQLite database (project.db) for structured supply chain data.
• LLM-powered SQL generation using ASI:ONE.
• Document-based similarity search for unstructured data.
• Hybrid reasoning that combines structured and unstructured data.
• Fallback logic for gracefully handling unknown or poorly classified queries.

FastAPI Application Summary

The backend is built with FastAPI and provides a secure, modular interface for document upload, vectorization, querying, and admin-controlled operations. Key highlights include:

• Document Upload & Processing: Uploaded PDF documents are stored, chunked, and embedded using Google Generative AI embeddings. The chunks are indexed using FAISS, and stored in a local vector database directory for retrieval.
• Document Management: A DocManager class handles all operations related to uploading, retrieving, listing, and deleting document metadata and vectorstore directories.
• Query Execution: The /query endpoint writes natural language questions to a file (query_to_agents.txt) to be picked up by backend agents. Responses are read from response_from_agents.txt and returned to the user.
• Authentication: Admin authentication is enabled via a login form, generating a session token to restrict document operations.
• SQLite Upload Support: A separate endpoint allows uploading .db files for SQL-based agents to query against.
• CORS & Static Hosting: CORS is fully enabled for dev flexibility. Static Vue.js frontend files are served from the /static directory.

Frontend: A lightweight Vue.js frontend is used, embedded directly via Vue CDN (https://unpkg.com/vue@3). This avoids build complexities and allows tight integration with the static file-based FastAPI hosting setup.

FastAPI Workflow Chart

Technologies Used

• Python – Core programming language for building the agent logic and backend.
• uAgents – Agent-based microservices framework for defining and managing autonomous agents.
• Fetch AI ASI:ONE API – Used for advanced natural language understanding and SQL generation via large language models.
• SQLite – Lightweight relational database for storing structured data such as orders, customers, and products.
• FAISS – Facebook AI Similarity Search library for fast and efficient vector-based semantic retrieval over PDF documents.

🛰️ Agent Addresses (Innovation Lab)

This project is part of the Fetch.ai Innovation Lab and includes the following autonomous agents:

• Main Agent (Intent Classifier Agent)
  test-agent://agent1qv68xelknm87mqkw3v8czfw3fnujdwzra09x8m32ff43lwzeefe67h9d675

• Document Agent
  test-agent://agent1qg4kpt4nmjckg9umkapx7v45mx8h02342mhn8g0aez5lzzx0mcj72a0ncfj

• Hybrid Agent
  test-agent://agent1qf05xu5emqz3y7gemadh08ekxnxvs6g9kj0ha068n8y3mzh94sn7cwrwlc8

• SQL Agent
  test-agent://agent1qw4z53dh3ttmdqku5q0xpqm6s25a0g69fsgh3lr440ak38stvfcy79fh2lk

• Worker Agent Address
  test-agent://agent1qv68xelknm87mqkw3v8czfw3fnujdwzra09x8m32ff43lwzeefe67h9d675

Screenshots

Answer from Agent Flow on ASI:ONE

Chat Interface

Admin Panel

Admin Login

Document Management

Document Upload

Database Upload

Database Connection Details

AgentVerse Integration

AgentVerse Agent Details

AgentVerse Local Agents

AgentVerse Intent Agent Chat

AgentVerse Logs

API Documentation

FastAPI Docs

Info Page

System Logs

Getting Started

1. Clone the Repository

git clone https://github.com/0rajnishk/AgentFlow.git
cd agent-flow

2. Set Environment Variables

Create a .env file or export the following environment variable:

export GOOGLE_API_KEY="your-ASI:ONE-api-key"

3. Install Dependencies

pip install -r requirements.txt

4. Run the System

Launch each agent independently or as a managed process depending on your architecture. Example:

python agents/sql_agent.py
python agents/hybrid_agent.py
python agents/document_agent.py
python agents/error_handler_agent.py

update the address of each of the above agents in the intent_classifier.py file

# Replace these with the actual addresses printed by each worker agent at startup
SQL_AGENT_ADDR = "agent1qw4z53dh3ttmdqku5q0xpqm6s25a0g69fsgh3lr440ak38stvfcy79fh2lk"
HYBRID_AGENT_ADDR = "agent1q0sh6f3n2r8azrs524chrn0e7h7p3qkm25v502jzczkrgjmtnhe972h2g64"
DOC_AGENT_ADDR = "agent1q2tpmsy506wtsdn0j7823s2vdm7f50l485azzc6z8lh2zk50cqwevn03e6q"
ERROR_AGENT_ADDR = "agent1q08hrn7j6t7ywmwdllrvl903t08sn4xd2ua2t4hy4kd6uxspqq0rgaudrpx"

python agents/intent_classifier_agent.py

Run them main.py our FastAPI server

uvicorn main:app --host 127.0.0.1 --port 8000 --reload

Example Query Flow

1. User asks: "Show me the top 5 most expensive products."
2. Intent Classifier routes to SQL Agent.
3. SQL Agent generates SQL using ASI:ONE, executes on project.db, and summarizes.
4. Response returned to the user through the intent classifier.