Human-in-the-Loop RAG Agent

Why it's ranked here

The project makes source control concrete instead of merely showing citations after generation. Users can reject weak context before synthesis, while semantic and hybrid retrieval remain configurable. That transparency comes with meaningful setup and an intentionally interrupted conversation flow.

What's good

The approval boundary is explicit: retrieval stops before answer synthesis, and the final response uses only selected chunks. Search controls include result limits, similarity thresholds, and semantic or hybrid matching. State snapshots also expose loading, errors, query history, and knowledge-base status.

Tradeoffs

Running it requires Python, Node.js, PostgreSQL with pgvector, and an OpenAI-compatible model and embedding endpoint. Each substantive question pauses for manual source selection. Frontend and backend state definitions must match exactly, creating a maintenance point when the shared contract changes.

How to use it well

Use it for internal knowledge workflows where a person should inspect evidence before an answer reaches others. It suits teams evaluating agent state synchronization and controlled retrieval. It does not serve workflows needing fully automatic answers without an approval pause.

Observed

Application shape

Full-stack browser application with a React frontend and Python backend.

Backend interface

FastAPI service carrying AG-UI traffic through Server-Sent Events, plus a health endpoint.

Install surface

Backend dependencies install with uv; frontend dependencies install with npm.

Runtime requirements

Python 3.12+, Node.js 18+, PostgreSQL with pgvector, and an OpenAI-compatible API endpoint.

Retrieval modes

Supports semantic vector search and hybrid vector plus text search.

Ingestion surface

Includes a command-line Python script that ingests Markdown documents into PostgreSQL.