How krbuti/setu-compliance-rag works
SETU Compliance Policy RAG answers policy questions through a Gradio or CLI entry point, classifies each query, then retrieves grounded policy information, runs policy-navigation tools when needed, and combines available results into a cited response. A separate preparation path converts policy PDFs to JSON and indexes chunks in persistent ChromaDB. The README documents the major retrieval, chunking, reflection, and provider capabilities; sampled source confirms key orchestration and indexing calls. Some retrieval internals and PDF conversion details are documented but their implementations were not sampled.
Main entry points: main_agent.py Query Orchestrator · llm_config.py Model Factories · information_agent.py Policy Retrieval · ingest.py Dataset Ingestion
Components 20
User Interface
- Gradio Chat
app_gradio.py
Query Answering
- Query Orchestrator
main_agent.py - Query Triage
triage_agent.py - Policy Retrieval
information_agent.py - Hybrid Search
information_agent.py - Reranking
information_agent.py - Answer and Reflection
information_agent.py - Policy Navigation Tools
action_agent.py - Response Combiner
main_agent.py
Policy Indexing
- Policy PDFs
- PDF Conversion
build_dataset.py - Policy JSON Dataset
dataset - Table-aware Chunking
data_preprocessing.py - Dataset Ingestion
ingest.py
Models and Storage
- Model Factories
config/llm_config.py - LLM Providers
- Embedding Provider
- ChromaDB Index
chroma_db
Other components
- Policy User
- Cited Response
Connections 24
- Policy User → Gradio Chat: submits query
README.md - Gradio Chat → Query Orchestrator: invokes
app_gradio.py - Policy User → Query Orchestrator: submits query
README.md - Query Orchestrator → Query Triage: classifies via
main_agent.py - Query Orchestrator → Policy Retrieval: routes to
main_agent.py - Query Orchestrator → Policy Navigation Tools: routes to
main_agent.py - Query Triage → Query Orchestrator: returns routing
main_agent.py - Policy Retrieval → Hybrid Search: retrieves with
README.md - Hybrid Search → ChromaDB Index: searches
README.md - Hybrid Search → Reranking: ranks results
README.md - Policy Retrieval → Answer and Reflection: generates and verifies
README.md - Policy Retrieval → Model Factories: gets model
information_agent.py - Policy Navigation Tools → Model Factories: gets model
action_agent.py - Query Orchestrator → Response Combiner: combines results
main_agent.py - Response Combiner → Model Factories: gets model
main_agent.py - Response Combiner → Cited Response: returns response
README.md - Policy PDFs → PDF Conversion: parsed into
README.md - PDF Conversion → Policy JSON Dataset: writes dataset
README.md - Policy JSON Dataset → Dataset Ingestion: loads documents
ingest.py - Dataset Ingestion → Table-aware Chunking: creates chunks
ingest.py - Dataset Ingestion → Model Factories: gets embeddings
ingest.py - Model Factories → LLM Providers: creates client
README.md - Model Factories → Embedding Provider: creates client
README.md - Dataset Ingestion → ChromaDB Index: writes chunks
ingest.py