Fix RAG retrieval silent failure and question key mismatch (#130)

[rahulkate173]

Fix for Defect 1 (Core Retrieval Bug)

• Renamed Variable: Changed self.retriever to self.vector_store to store the raw FAISS index instead of the LangChain VectorStoreRetriever wrapper.
• Updated setup_vectorstore(): Removed the .as_retriever() call and now assigns the FAISS object directly to self.vector_store.
• Rewrote get_relevant_document():
  • Added a None guard to handle cases where the vector store hasn't been initialized.
  • Now calls self.vector_store.similarity_search_with_score(query, k=1) directly to retrieve (Document, float) tuples (where the float is the L2 distance).
  • Converted the L2 distance to a similarity score using 1.0 / (1.0 + distance). This ensures the threshold comparison works as expected (higher score = more similar).

Note Regarding Defect 2

The reported wrong-key bug ("query" vs "question") appears to have been resolved in a prior refactor. The run() method no longer uses a LangChain chain with RunnablePassthrough. Instead, it already formats the prompt directly via self.prompt_template.format(question=..., context=...) and calls self.provider.generate(). No further changes were required for this defect.

Closes #130

[Noaman-Akhtar]

i went through both PR 160 and PR 161, and they look like two different solutions for the same current retrieval issue. am i missing some difference between them, or is PR 161 meant to replace PR 160?

[rahulkate173]

Yes pr 161 is meant to replace pr 160

[Noaman-Akhtar]

Then you should close pr 160 , so it doesn't create confusion

[rahulkate173]

Done

Pr 160 closed