How kchimbodza/agentic-researcher works
Agentic Researcher is a desktop-style PDF research app: a user asks a multi-step question, agents retrieve and verify evidence, and a reviewer can approve a partial answer and sign off. Its main runtime path is exposed through the web server and LangGraph agent; retrieval is served through MCP, with a direct-call fallback. The README also documents evaluation and a fine-tuned claim checker. Some source is unsampled, so the graph omits unverified connections rather than inferring them.
Main entry points: agent.py Critic · server.py Web server · tools.py Search tools · evaluate.py Evaluation runner
Components 23
Research workflow
- Agent graph
app/agent.py - Planner
app/agent.py - Researcher
app/agent.py - Critic
app/agent.py - Writer
app/agent.py - Model router
app/llm.py
Documents and retrieval
- Tool client
app/mcp_client.py - MCP tool server
app/mcp_server.py - Search tools
app/tools.py - Vector index
app/index.py - PDF library
app/docs.py
App and review
- Browser UI
app/static/app.js - Web server
app/server.py - Audit and threads
app/store.py
Evaluation and checking
- Claim checker
app/claimcheck.py - Evidence guardrails
app/guardrails.py - Evaluation runner
evals/evaluate.py - RAG baseline
evals/baseline.py - Answer scoring
evals/scoring.py - Checker training
finetune/train.py
Other components
- Questioner
- Model providers
- Reviewer
Connections 29
- Questioner → Browser UI: asks via
README.md - Browser UI → Web server: submits question
- Web server → Agent graph: runs
app/server.py - Agent graph → Planner: plans lookups
README.md - Planner → Researcher: dispatches lookups
README.md - Researcher → Tool client: searches via
README.md - Tool client → MCP tool server: calls tools
app/mcp_client.py - Tool client → Search tools: falls back to
app/mcp_client.py - MCP tool server → Search tools: serves tools
app/mcp_server.py - Search tools → Vector index: vector searches
app/tools.py - Search tools → PDF library: keyword searches
app/tools.py - Vector index → PDF library: indexes pages from
app/index.py - Researcher → Model router: extracts claims with
README.md - Critic → Search tools: checks page quotes
app/agent.py - Critic → PDF library: reads page text
app/agent.py - Critic → Evidence guardrails: checks claims with
app/agent.py - Critic → Claim checker: checks support with
app/agent.py - Critic → Researcher: retries lookup
README.md - Critic → Writer: passes verified evidence
README.md - Writer → Model router: writes with
README.md - Model router → Model providers: routes requests
app/llm.py - Web server → Audit and threads: records run and audit
app/server.py - Web server → Audit and threads: loads thread history
app/server.py - Reviewer → Web server: approves or reviews
app/server.py - Web server → Browser UI: streams events
- Evaluation runner → Agent graph: evaluates
evals/evaluate.py - Evaluation runner → RAG baseline: runs baseline
evals/evaluate.py - Evaluation runner → Answer scoring: scores answers
evals/evaluate.py - Checker training → Claim checker: produces checker model