How hsmshubham007-code/week5task1 works
NovaTech is a policy-document RAG assistant: users ask questions in a Streamlit chat and receive evidence-grounded answers or refusals with source/page citations. Its data path extracts PDF pages, creates differently sized chunks, and builds Chroma collections; the QA runtime retrieves chunks, selects and scores evidence, then answers or refuses. Separate evaluation scripts compare baseline, hybrid, and reranked retrieval and test generation. The README documents capabilities beyond the sampled runtime excerpts; links based on those capabilities cite the README.
Main entry points: golden_set.csv Golden questions · chroma Vector collections · grounded_qa.py Grounded QA runtime · app.py Streamlit chat
Components 19
Corpus ingestion
- Policy PDFs
data/documents - PDF extraction
src/ingestion/extract_pdfs.py - Extracted pages
data/processed - Token chunking
src/ingestion/chunk_documents.py - Chunk datasets
data/chunks - Chroma builder
src/ingestion/build_chroma.py
Grounded QA
- Vector collections
storage/chroma - Grounded QA runtime
src/generation/grounded_qa.py - Top-five retrieval
src/generation/grounded_qa.py - Evidence selection
src/generation/grounded_qa.py - Answer and citation
src/generation/grounded_qa.py
Evaluation
- Golden questions
data/golden_set/golden_set.csv - Recall evaluation
evals/evaluate_recall.py - BM25 hybrid evaluation
evals/evaluate_hybrid.py - Reranker evaluation
evals/evaluate_reranker.py - Generation evaluation
evals/evaluate_generation.py - Full evaluation
evals/evaluate_all.py
Other components
- Policy questioner
- Streamlit chat
app.py
Connections 23
- Policy questioner → Streamlit chat: asks
README.md - Streamlit chat → Grounded QA runtime: calls
app.py - Policy PDFs → PDF extraction: provides PDFs
src/ingestion/extract_pdfs.py - PDF extraction → Extracted pages: writes pages
src/ingestion/extract_pdfs.py - Extracted pages → Token chunking: reads pages
src/ingestion/chunk_documents.py - Token chunking → Chunk datasets: writes chunks
src/ingestion/chunk_documents.py - Chunk datasets → Chroma builder: reads chunks
src/ingestion/build_chroma.py - Chroma builder → Vector collections: builds collections
src/ingestion/build_chroma.py - Grounded QA runtime → Top-five retrieval: runs
README.md - Top-five retrieval → Vector collections: queries
README.md - Top-five retrieval → Evidence selection: supplies chunks
README.md - Evidence selection → Answer and citation: grounds or refuses
README.md - Answer and citation → Streamlit chat: returns result
app.py - Golden questions → Recall evaluation: evaluated against
evals/evaluate_recall.py - Golden questions → BM25 hybrid evaluation: evaluated against
evals/evaluate_hybrid.py - Golden questions → Reranker evaluation: evaluated against
evals/evaluate_reranker.py - Golden questions → Generation evaluation: evaluated against
README.md - Golden questions → Full evaluation: evaluated against
evals/evaluate_all.py - Recall evaluation → Vector collections: queries
evals/evaluate_recall.py - BM25 hybrid evaluation → Vector collections: queries vectors
evals/evaluate_hybrid.py - Reranker evaluation → Vector collections: retrieves candidates
evals/evaluate_reranker.py - Generation evaluation → Grounded QA runtime: evaluates answers
README.md - Full evaluation → Grounded QA runtime: evaluates generation
README.md