Project case study · Agentic AI · RAG · Evaluated system
Multi-agent AI research assistant with LangGraph and RAG
A full-stack research workflow that combines planning, retrieval, parallel research, tool execution, cited synthesis, bounded reflection, persistent state, and server-sent event streaming.
01 · Context
Research question
Research tasks often require several different operations: break down a query, retrieve relevant material, gather evidence, synthesize findings, and preserve progress. This project explores how to orchestrate those steps in a testable, stateful application.
02 · Methods
Approach
- Used LangGraph to coordinate five roles and parallel research branches.
- Connected ChromaDB retrieval with local ONNX embeddings and configurable research tools.
- Implemented FastAPI REST and server-sent event (SSE) interfaces, with SQLite checkpointing for persistent workflow state.
- Added citation-grounded synthesis, bounded reflection, LangSmith evaluation tooling, and deterministic offline tests.
03 · Evaluation
Reported results
- The repository reports 103 offline tests for application behavior in deterministic fake-LLM mode.
- The workflow exposes REST and SSE interfaces and includes five agent roles.
- Tests exercise software contracts; they are not a measure of answer correctness, evidence quality, or research usefulness.
04 · Scope
Limitations
The deterministic fake-LLM test mode validates application behavior rather than live-model answer quality. The system has not been shown to outperform human research workflows or establish the correctness and completeness of generated research. Human verification of cited sources remains necessary.