sruthisree11/ncert-q-a-bot

Purpose and Entry Points

This is a local, CLI-based RAG assistant for answering questions over NCERT textbook PDFs. It has two operational entry points:

• Offline corpus ingestion/index construction: src/build_vector_store.py

• Interactive retrieval and answer generation: src/qa_cli.py

Source Corpus and Persistent Knowledge Store

PDFs are the authoritative document corpus under data/pdfs/, currently including NCERT files such as data/pdfs/hesc106.pdf. The build stage converts this corpus into two coupled local artifacts:

• Serialized chunk metadata/text: data/chunks.pkl

• FAISS nearest-neighbor index: data/chunks_index.faiss

The index and pickle must remain aligned: FAISS retrieves vector positions, while the pickle provides the corresponding textual context.

Ingestion and Index-Build Pipeline

src/build_vector_store.py is the batch pipeline boundary between raw documents and retrievable knowledge. It:

1. Reads PDFs from data/pdfs/.

2. Extracts document text via a PDF library.

3. Splits text into semantic chunks.

4. Creates embeddings.

5. Persists chunks and the FAISS index to data/chunks.pkl and data/chunks_index.faiss.

Rebuilding the store is required when PDFs or chunking/embedding behavior changes.

Retrieval and Answering Pipeline

src/qa_cli.py is the runtime application. It accepts a natural-language question from the terminal, loads the local FAISS index and chunk collection, embeds the question, and retrieves the most relevant chunks. Retrieved passages become the grounding context for Gemini rather than being answered solely from the model’s general knowledge.

The control flow is: user question → query embedding → FAISS similarity search → retrieved chunk context → prompt assembly → Gemini response → terminal output.

Prompt Boundary

prompts/qa_prompt.txt defines the answer-generation prompt template. It is the principal policy/context boundary between deterministic local retrieval and the external generative model. Changes here can alter how retrieved textbook excerpts are presented, how strictly answers follow context, and the final response format without rebuilding the vector store.

External Services and Runtime Dependencies

The system runs on Python 3 with local FAISS CPU search and local persisted artifacts. Google Gemini is the external generation service, authenticated through GEMINI_API_KEY loaded from a root .env file. Relevant dependency expectations are declared in requirements.txt, including FAISS, PDF extraction, dotenv, and Google’s generative AI SDK.

Architectural Constraints

The repository is intentionally single-machine and file-backed: no web server, database, queue, or remote vector database is present. Index construction and question answering are separate commands, allowing query-time execution to avoid PDF parsing and index generation. Data quality depends on readable PDFs and successful extraction from data/pdfs/.