Enterprise RAG Knowledge Assistant

Users need a practical way to ask questions over documents without sending the entire document collection into every model request.

The application loads supported documents, splits them into overlapping chunks, creates normalized BGE-small-en-v1.5 embeddings on CPU, stores them in FAISS, retrieves the top four matches, and supplies the built context to Gemini 2.5 Flash.

• PDF, DOCX, TXT, Markdown, and CSV ingestion
• Chunk size 1000
• Chunk overlap 200
• CPU embeddings
• Normalized embeddings
• Top-k retrieval of 4
• Thread-based state

Knowledge spread across PDFs, DOCX files, text, Markdown, and CSV needs a focused retrieval workflow before an LLM can answer questions over it.

The assistant turns uploaded documents into searchable chunks, retrieves the most similar context, and generates an answer using that context.

The loader supports PDF, DOCX, TXT, Markdown, and CSV documents.

Documents use a chunk size of 1000 with 200-token overlap. BAAI/bge-small-en-v1.5 creates normalized CPU embeddings.

FAISS retrieves the top four similar chunks. The context builder passes those results to Gemini 2.5 Flash at temperature 0.3.

The workflow is organized as retrieve, build_context, and generate_answer, with thread-based state.

The implementation makes the ingestion, retrieval, and generation stages explicit so the grounded-answer path remains easy to inspect.