Fix RAG Retrieval Quality pipeline & Add Source Citations to AI Document Assistant

We have an existing RAG-based document Q&A prototype, but the retrieval quality is inconsistent. The assistant can answer questions, but in some cases the retrieval layer returns loosely related chunks instead of the most relevant sections of the document. This leads to incomplete or inaccurate answers even when the correct information exists in the uploaded documents. We need an AI engineer to review and improve the retrieval pipeline and add reliable source references to generated answers. Scope: • Review the existing RAG retrieval flow and identify the cause of poor retrieval • Evaluate chunking strategy, embedding queries, top-k retrieval and similarity thresholds • Improve vector search so relevant document sections rank higher • Reduce irrelevant context being passed to the LLM • Add source/document references to answers using retrieved chunk metadata • Test retrieval quality against a set of documents we provide and sample questions The document ingestion and LLM integration are already implemented. This is a focused debugging and optimization task rather than a complete RAG application build. Experience with RAG, embeddings, vector databases, Python and LLM APIs preferred.