k-electron/cfr-copilot

How k-electron/cfr-copilot works

CFR Copilot is a local FDA regulatory-compliance assistant. Users ask regulatory questions through the CLI or Streamlit chat; the agent can choose regulation search or structured recall queries, while plain and structured RAG offer direct retrieval and citation-checked answers. Ingestion loads eCFR sections and openFDA recalls into separate search and relational stores. The README also documents retrieval/answer evaluation and optional tracing. Edges use README statements or SOURCE INDEX references; unsampled subsystems are represented only where those references support them.

Main entry points: cli.py CLI · agent.py Tool-using agent · index.py CFR indexer · retriever.py Composed retriever

Components 28

User interfaces

CLI src/cfr_copilot/cli.py
Chat UI app/streamlit_app.py

Question answering

Tool-using agent src/cfr_copilot/agent/agent.py
Retrieval context src/cfr_copilot/agent/context.py
Agent tools src/cfr_copilot/agent/tools.py
Agent middleware src/cfr_copilot/agent/middleware.py
RAG chain src/cfr_copilot/chains/rag.py
Verified answers src/cfr_copilot/chains/structured.py
Answer schemas src/cfr_copilot/chains/schemas.py
Composed retriever src/cfr_copilot/retrieval/retriever.py
Reranking src/cfr_copilot/retrieval/rerank.py
Local model client src/cfr_copilot/llm.py

Data ingestion

eCFR loader src/cfr_copilot/ingest/ecfr.py
Legal chunking src/cfr_copilot/ingest/chunking.py
CFR indexer src/cfr_copilot/ingest/index.py
Recall loader src/cfr_copilot/ingest/openfda.py

Search and storage

Hybrid vector store src/cfr_copilot/retrieval/vectorstore.py
Full CFR sections src/cfr_copilot/sections.py
Postgres pool src/cfr_copilot/db.py
Configuration src/cfr_copilot/config.py

Evaluation and observability

Evaluation dataset src/cfr_copilot/evaluation/dataset.py
Retrieval evaluation src/cfr_copilot/evaluation/retrieval.py
Answer evaluation src/cfr_copilot/evaluation/answers.py
Tracing src/cfr_copilot/tracing.py

Other components

User
eCFR service
openFDA service
Ollama

Connections 37

User → CLI: asks via CLI README.md
User → Chat UI: asks in chat README.md
CLI → Tool-using agent: runs agent src/cfr_copilot/cli.py
Chat UI → Tool-using agent: streams agent README.md
Tool-using agent → Retrieval context: uses context src/cfr_copilot/agent/agent.py
Tool-using agent → Agent tools: dispatches tools src/cfr_copilot/agent/agent.py
Tool-using agent → Agent middleware: installs middleware src/cfr_copilot/agent/agent.py
Agent tools → Composed retriever: searches regulations src/cfr_copilot/agent/tools.py
Agent tools → Postgres pool: queries recalls src/cfr_copilot/agent/tools.py
CLI → RAG chain: runs direct RAG src/cfr_copilot/cli.py
CLI → Verified answers: requests verified answer src/cfr_copilot/cli.py
Verified answers → RAG chain: builds on RAG src/cfr_copilot/chains/structured.py
Verified answers → Answer schemas: returns typed answer src/cfr_copilot/chains/structured.py
RAG chain → Hybrid vector store: retrieves documents src/cfr_copilot/chains/rag.py
Composed retriever → Hybrid vector store: searches candidates src/cfr_copilot/retrieval/retriever.py
Composed retriever → Reranking: reranks candidates src/cfr_copilot/retrieval/retriever.py
Composed retriever → Full CFR sections: expands sections src/cfr_copilot/retrieval/retriever.py
RAG chain → Local model client: generates answer src/cfr_copilot/chains/rag.py
Tool-using agent → Local model client: creates agent model src/cfr_copilot/agent/agent.py
eCFR service → eCFR loader: provides CFR XML src/cfr_copilot/ingest/ecfr.py
eCFR loader → Legal chunking: supplies sections src/cfr_copilot/ingest/chunking.py
CLI → CFR indexer: runs indexing src/cfr_copilot/cli.py
CFR indexer → eCFR loader: loads CFR src/cfr_copilot/ingest/index.py
CFR indexer → Legal chunking: chunks sections src/cfr_copilot/ingest/index.py
CFR indexer → Hybrid vector store: writes vectors src/cfr_copilot/ingest/index.py
CFR indexer → Full CFR sections: stores full sections src/cfr_copilot/ingest/index.py
openFDA service → Recall loader: provides recall data src/cfr_copilot/ingest/openfda.py
Recall loader → Postgres pool: stores recalls src/cfr_copilot/ingest/openfda.py
Full CFR sections → Postgres pool: reads and writes src/cfr_copilot/sections.py
Hybrid vector store → Local model client: embeds queries src/cfr_copilot/retrieval/vectorstore.py
Local model client → Ollama: calls local models src/cfr_copilot/llm.py
CLI → Evaluation dataset: loads evaluation set src/cfr_copilot/cli.py
Retrieval evaluation → Evaluation dataset: uses questions src/cfr_copilot/evaluation/retrieval.py
Retrieval evaluation → Composed retriever: measures retrieval src/cfr_copilot/evaluation/retrieval.py
Answer evaluation → Evaluation dataset: uses questions src/cfr_copilot/evaluation/answers.py
Answer evaluation → Verified answers: evaluates answers src/cfr_copilot/evaluation/answers.py
Chat UI → Tracing: optional tracing

Drawn by GitDiagram on November 12, 2026 from the repository's file tree, README and source files. Read as Markdown ·