ResearchFlow AI is a production-grade agentic research system that decomposes complex questions into sub-queries, retrieves and verifies web evidence, and synthesizes citation-grounded answers — with self-correcting retry logic when evidence quality is low.
Ask any complex research question — ResearchFlow decomposes it, searches for evidence, verifies quality, and returns a cited answer with a full reasoning trace.
Most LLM apps do: question → retrieve → answer
ResearchFlow does:
question → decompose → search → verify confidence →
if low: re-search with refined query (cyclic retry) →
answer with citations + reasoning trace
This is agentic reasoning — the system evaluates its own evidence quality and decides whether to retry before synthesizing. Single-prompt LLM systems hallucinate. ResearchFlow verifies.
graph TD
A[User Query] --> B[Planner Node]
B --> C[Decomposer Node]
C --> D[Search Node - Tavily]
D --> E[Verifier Node]
E -->|confidence < threshold| D
E -->|confidence ≥ threshold| F[Synthesizer Node]
F --> G[Citation-Grounded Answer]
| Node | Role |
|---|---|
| Planner | Analyzes query intent, defines research strategy |
| Decomposer | Breaks complex question into focused sub-queries |
| Search | Retrieves web evidence via Tavily Search API |
| Verifier | Scores evidence relevance and confidence (0–1) |
| Synthesizer | Generates final answer with inline citations |
The cyclic edge between Verifier and Search is the core innovation — if confidence falls below threshold, the system refines the query and searches again rather than hallucinating from weak evidence.
Query: "How does LangGraph enable self-correcting AI agents?"
Pipeline trace:
[Planner] → Strategy: technical documentation + recent examples
[Decomposer] → Sub-queries: ["LangGraph cyclic workflow",
"LangGraph conditional routing",
"LangGraph stateful agents retry"]
[Search] → Retrieved 9 sources via Tavily
[Verifier] → Confidence: 0.44 (below threshold) → RETRY
[Search] → Refined query: "LangGraph self-correction checkpointing 2024"
[Verifier] → Confidence: 0.83 → PASS
[Synthesizer] → Answer with 5 inline citations
Sample output:
• LangGraph enables self-correcting agents through cyclical workflows where
an agent generates output, evaluates results, and loops back for revision.
• It maintains shared state and checkpoints so intermediate results persist
across iterations.
• Conditional routing allows workflows to retry, revise, or terminate
execution based on intermediate outcomes.
Limitation: LangGraph provides orchestration infrastructure, but evaluation
logic must still be implemented by the developer.
- Python 3.10+
- Ollama running locally with Llama 3.2:
ollama pull llama3.2 - Tavily API key (free tier at tavily.com)
git clone https://github.com/krishnakoushik225/langgraph-research-agent
cd langgraph-research-agent
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txtcp .env.example .env
# Add your keys:
# TAVILY_API_KEY=your_key_here
# LANGSMITH_API_KEY=your_key_here (optional, for tracing)# Backend
uvicorn app.main:app --reload
# Runs at http://127.0.0.1:8000
# Frontend (separate terminal)
cd frontend && npm install && npm start
# Runs at http://localhost:3000| Layer | Technology |
|---|---|
| Agent Orchestration | LangGraph (cyclic state machine) |
| LLM Backend | Ollama + Llama 3.2 (local, private) |
| Web Search | Tavily Search API |
| API Backend | FastAPI (async) |
| Frontend | React + Axios |
| Observability | LangSmith (optional tracing) |
langgraph-research-agent/
├── backend/
│ └── app/
│ ├── graph/
│ │ ├── planner.py # Research strategy generation
│ │ ├── decomposer.py # Sub-question generation
│ │ ├── search.py # Tavily evidence retrieval
│ │ ├── verifier.py # Confidence scoring + retry logic
│ │ └── synthesizer.py # Citation-grounded answer generation
│ ├── services/
│ │ └── ollama_client.py # Local LLM client
│ └── main.py # FastAPI entrypoint
├── frontend/
│ └── src/
│ ├── App.jsx
│ └── App.css
├── docs/
│ ├── researchflow-demo.png
│ └── architecture.png
├── .env.example
├── requirements.txt
└── README.md
Evidence Quality — Web search returns noisy results. The Verifier node assigns a confidence score and triggers re-search with refined queries when quality is low.
Hallucination Prevention — The Synthesizer is constrained to generate answers only from retrieved evidence, with every claim linked to a source.
Agent Coordination — LangGraph manages shared workflow state across all nodes, enabling conditional branching and retry without losing intermediate results.
Adaptive Research — If initial sub-questions yield weak evidence, the system reformulates them before attempting synthesis.
LangGraph was chosen specifically because it supports cyclic graphs — standard LangChain chains are DAGs (directed acyclic graphs) and cannot model retry loops. LangGraph enables:
- Stateful workflows — shared research state persists across all nodes
- Conditional routing — retry, revise, or terminate based on verifier output
- Modular reasoning nodes — each node has a single, testable responsibility
- Checkpointing — intermediate results survive failures and restarts
- LangSmith full trace dashboard
- Human-in-the-loop checkpoint before synthesis
- Vector database retrieval (Pinecone) as additional search source
- Streaming tokens to frontend in real-time
- Multi-model support (GPT-4, Claude, Gemini)
- Multi-agent collaboration mode
MIT — free to use and build on.
Built by Krishna Koushik Unnam · AI Systems Developer