RagFlowProMax

A multi agent enterprise RAG. A supervisor routes specialist agents and a verifier checks the answer. Part of the RagFlow line.

Part of the RagFlow line, a series of reference enterprise RAG implementations. This repository is RagFlowProMax, Multi agent enterprise RAG. See the full line below.

RagFlowProMax does not rely on a single chain or a single agent. A supervisor plans the work and routes the question to specialist worker agents, a synthesizer merges their findings, and a verifier checks the answer is grounded before it is returned. The document worker is the RagFlowProPlus self correcting RAG over pgvector, so the system is an agent of agents. It runs fully locally on Ollama, with the Claude 5 family as the frontier cloud option.

The clip above is a live, unedited run on a local qwen2.5 model over pgvector. The expandable trace shows the supervisor route the question, the document agent answer, and the verifier check it. No paid keys were used.

The agents

Agent	Role
Supervisor	Plans and routes the question to the right specialist agents, bounded so it always terminates
Document agent	The RagFlowProPlus self correcting RAG over pgvector: retrieve, grade, rewrite, generate, self check
Web agent	Grounded web search, off by default, used only when the supervisor asks for it
Synthesizer	Merges the specialist findings into one grounded answer with sources
Verifier	Checks the answer is grounded in the findings before it is returned

Every agent's contribution is recorded in a trace returned with the answer, the observability spine from the enterprise design. The orchestration is bounded, which is the cost guard.

Architecture

graph TD
    User[User] --> API[FastAPI backend]
    API --> Supervisor[Supervisor agent, plan and route]
    Supervisor --> DocAgent[Document agent, self correcting RAG]
    Supervisor --> WebAgent[Web agent, optional]
    DocAgent --> PG[(Postgres with pgvector)]
    DocAgent --> Synth[Synthesizer]
    WebAgent --> Synth
    Synth --> Verifier[Verifier, grounded check]
    Verifier --> Answer[Answer plus agent trace]

Loading

How to use

Local, fully offline with Ollama (no paid keys)

# 1. Data services
make db-up             # postgres with pgvector, plus redis

# 2. Ollama and the local models
ollama serve &
ollama pull nomic-embed-text
ollama pull qwen2.5:7b-instruct

# 3. Install and run
make install
EMBEDDING_PROVIDER=ollama make dev        # API on :8000
make frontend                             # UI on :8501, second terminal

Load the bundled sample data with make load-samples, then ask a question and open the trace to watch which agents ran.

Configuration

Setting	Default	Meaning
EMBEDDING_PROVIDER	google	google or ollama
AGENT_ENABLE_WEB	false	grounded first; turn on to let the supervisor use the web agent
AGENT_CONFIDENCE_THRESHOLD	0.6	document agent grade gate
AGENT_MAX_STEPS	12	hard cap on the document agent's internal steps
API_KEY	change_me	required in the X-API-Key header

API reference

Method and path	Purpose
GET /health	Liveness, no auth
POST /v1/chat	Multi agent answer with the agent trace and which agents ran
POST /v1/upload-doc	Upload and asynchronously index a document
GET /v1/list-docs	List indexed documents
POST /v1/delete-doc	Delete a document and its chunks
GET /metrics	Prometheus metrics

Testing

make test        # unit tests, no database or model needed

The RagFlow line

RagFlowProMax is one implementation in the RagFlow line, a series demonstrating distinct enterprise RAG retrieval strategies. The whole line is measured on the same golden set in the rag-catalog benchmark.

Year	Repository	Generation
2022	RagFlow	Naive RAG, single dense retrieval
2023	RagFlowPlus	Advanced RAG, hybrid retrieval and reranking
2024	RagFlowPro	Modular production RAG, pgvector, streaming, evaluation
2025	RagFlowProPlus	Agentic RAG, self correcting with confidence grading
2026	RagFlowProMax, this repo	Multi agent enterprise, supervisor and specialist agents

Author

Malav Patel. GitHub @mlvpatel.

License

Released under the MIT License. See LICENSE. MIT is the simplest and most permissive of the common licenses, so anyone can read, run, modify, and reuse the code freely.