Complete agentic RAG architecture and documentation

[tusharg007]

Summary

• complete the production-oriented agentic RAG runtime hardening and retrieval-quality pass
• add repository documentation covering architecture, agentic behavior, reproducibility, testing, file mapping, troubleshooting, and evaluation limits
• add GitHub Actions CI for compile checks, pytest, and offline smoke evaluation

Architecture

• real LangGraph-based orchestration with deterministic fallback compatibility
• scoped retrieval across uploaded and sample corpora
• bounded prompt compaction, grounding verification, and single-pass repair flow
• hardened ingestion, upload validation, health checks, and graceful provider-error handling

Key capabilities

• cross-corpus RFP analysis with separated target and case-study evidence
• deterministic and idempotent ingestion with duplicate-file and duplicate-chunk protection
• grouped source traces, conversational follow-up handling, and safer insufficient-evidence behavior
• offline smoke evaluation plus real knowledge-base evaluation

Testing

• python -m py_compile app.py agent.py rag_engine.py config.py document_generator.py
• python -m compileall -f src tests evals
• python -m pytest -q

Evaluation

• python -m evals.run_evals
• python -m evals.run_kb_evals
• real KB evaluation runs against generated sample PDFs, a temporary evaluation upload fixture, and an isolated temporary vectorstore

Documentation

• rewrite README to match the current src-package architecture and runtime behavior
• add dedicated architecture, agentic RAG, reproducibility, testing/evaluation, file-map, and troubleshooting docs
• document deployment boundaries, safety limits, and known limitations without overstating production readiness

Safety and repository hygiene

• add CI workflow badge target and workflow file
• harden .gitignore for secrets, uploads, eval outputs, vectorstores, logs, caches, and local environments
• remove tracked egg-info metadata from the repository

Known limitations

• Streamlit deployment remains a single-user demonstration architecture
• local vectorstore persistence and synthetic sample PDFs are still part of the current design
• automated grounding and evaluation improve safety but do not prove universal correctness