AI Research Agent V1.0

A production-oriented autonomous research agent that combines web research, local knowledge retrieval, multi-step planning, reflection, and rigorous citation validation.

Why this project?

Modern LLMs hallucinate. They confidently state things that are factually wrong, and rarely distinguish between what they know and what they are inventing. This project explores a pragmatic mitigation: force the model to produce citations from explicitly retrieved sources, and automatically verify those citations before the answer is returned.

This implementation deliberately avoids black-box RAG frameworks in favour of a fully traceable, testable pipeline where every grounding decision is observable.

---

What happens when a user asks a question?

The agent executes a dynamic, multi-step pipeline. Simple questions take a fast path, while complex research triggers the full autonomous loop:

User Query
   ↓
Research Planning
   ↓
Tool Selection (LLM chooses actions)
   ↓
Web / Local Retrieval (DuckDuckGo / ChromaDB)
   ↓
Evidence Collection
   ↓
Reflection (Is this enough to answer?)
   ↓
Evidence Synthesis
   ↓
Claim Assessment (Fact-checking vs Evidence)
   ↓
Grounding Check (Rewriting unsupported claims)
   ↓
Citation Validation
   ↓
Final Answer

---

Technical Highlights

This project demonstrates several advanced AI engineering patterns:

• LangGraph-based Orchestration: A stateful StateGraph for the autonomous research loop, replacing brittle prompt-chaining.
• Multi-Step Research Loop: The agent can iteratively search, collect evidence, and reflect multiple times before synthesising.
• Dual-Channel Retrieval: Fuses live web search (ddgs) with a local RAG vector store.
• Reflection & Sufficiency: An LLM-as-a-judge node evaluates if the gathered evidence actually answers the prompt.
• Grounding & Citation Validation: Extracts factual claims, assesses them against retrieved evidence via token-overlap, and strictly validates all citation tags (e.g., [WEB: ...]).
• Provider Abstraction: Seamlessly switch between Gemini, OpenRouter, and Groq via environment variables.
• Session Memory: UUID-keyed conversation tracking.
• Containerised Deployment: Dockerized setup with build-time RAG ingestion, deployed on Render.
• Comprehensive Testing: 125 offline unit tests covering routing, fallback mechanisms, and quality checks.

---

Demo

Demo

The production deployment of AI Research Agent V1.0:

Live demo: https://ai-research-agent-sj1r.onrender.com/

---

Documentation Hub

Detailed documentation is available in the docs/ directory:

• 🏗️ Architecture — System diagram, LangGraph nodes, and core flow.
• 🔍 RAG Pipeline — Ingestion, chunking strategy, and multi-backend vector retrieval.
• 🛡️ Grounding & Citations — Claim extraction, conflict detection, and the grounding gate.
• 🔌 API Reference — FastAPI endpoints, request/response schemas, and trace payloads.
• 🚀 Deployment — Docker setup, build-time ingestion, and Render configuration.
• 🔧 Troubleshooting — Post-mortems for real issues fixed during V1.0 development.

---

Tech Stack

AI / Agent

• LangGraph: Stateful agent orchestration
• LangChain Core: Prompt templating

RAG / Retrieval

• ChromaDB: Primary local vector store
• Pinecone: Optional cloud fallback vector store
• Sentence-Transformers: all-MiniLM-L6-v2 local CPU embeddings
• DuckDuckGo Search: Live web retrieval (ddgs)

Backend

• Python 3.12
• FastAPI + Uvicorn: REST API and static file serving
• Pydantic: Configuration and schema validation

Frontend

• HTML/CSS/JS: Vanilla, zero-build-step frontend

Deployment

• Docker: Containerization with baked-in knowledge base
• Render: Web service hosting

Testing

• Unittest / HTTPX: 125 offline tests

---

Honest Limitations

This is a V1.0 release. It is not an AGI, nor is it 100% hallucination-free. The system has known constraints:

• Token-overlap grounding: Claim assessment relies on keyword overlap. Claims sharing keywords with semantically different evidence may pass incorrectly (semantic verification is planned).
• Web source quality: DuckDuckGo results are accepted without domain authority filtering.
• In-memory sessions: Sessions are lost on server restart (the frontend handles recovery gracefully).
• Synchronous execution: Long research queries hold the HTTP handler until complete.
• Provider rate limits: Extended tool loops may hit 429 limits on free-tier LLM providers.

---

Installation & Running Locally

Requirements: Python 3.12, pip

# 1. Virtual environment setup
python -m venv venv
.\venv\Scripts\Activate.ps1  # Windows
# source venv/bin/activate   # macOS / Linux

# 2. Install CPU-only PyTorch (required before other dependencies)
pip install torch --index-url https://download.pytorch.org/whl/cpu

# 3. Install remaining dependencies
pip install -r requirements.txt

# 4. Environment variables
cp .env.example .env
# Edit .env with your API keys (e.g., GEMINI_API_KEY)

# 5. Run the server
uvicorn api_server:app --host 127.0.0.1 --port 8000 --reload

Access the application at http://127.0.0.1:8000.

---

Docker

The Docker build automatically ingests the local knowledge base (docs/*.txt).

docker build -t ai-research-agent .

docker run -p 8000:8000 \
  -e GEMINI_API_KEY=your_key \
  -e LLM_PROVIDER=gemini \
  ai-research-agent

---

Testing

The project is heavily tested to ensure pipeline reliability without requiring network access.

python -m unittest discover

(Expect 125 tests to pass in ~2-4 seconds)

---

Project Status

V1.0 — Feature Complete and Frozen

Production deployment is live on Render. The core architecture is stable and thoroughly tested.

Author

Chandra Shekar
GitHub: chandra8688