DoRecifeChatFlow

Companion to do_recife_embedder, which builds the MongoDB Atlas vector store this flow reads from.

A minimal crewAI Flow wrapping a single conversational RAG agent that answers questions about the Diário Oficial do Recife, grounding every answer in the embeddings produced by do_recife_embedder.

The agent is transparent about tool use, stays strictly on the topic of the Diário Oficial, and always replies in the same language as the user.

Structure

• main.py – DoRecifeChatFlow, a Flow[Conversation] that appends the user message, calls the agent, and stores the reply.
• agents/ – DoRecifeAgent, a streaming LiteAgent (agent.kickoff()).
• tools/ – DoRecifeVectorSearchTool, a pre-configured MongoDBVectorSearchTool.
• types/ – Conversation and Message state models.

Requirements

• Python >=3.10, <3.14 and uv
• Access to the same MongoDB Atlas cluster the embedder wrote to
• An Anthropic API key (agent) and an OpenAI API key (query embeddings)

Setup

Install dependencies:

crewai install

Add the required variables to .env:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...
MONGODB_CONNECTION_STRING=mongodb+srv://...

ANTHROPIC_API_KEY runs the agent (anthropic/claude-haiku-4-5). OPENAI_API_KEY embeds the query (text-embedding-3-large, 3072-dim) and must match the embedder's settings or vector search returns irrelevant results.

Usage

crewai run

Change the question by editing the user_message input in src/do_recife_chat_flow/main.py.