Problem

The current HybridRetriever in csv_chroma.py retrieves documents
from multiple subdirectories using BM25 + SelfQuery + MultiQuery
expansion, resulting in ~90 documents being passed to
create_stuff_documents_chain.

All retrieved documents are stuffed directly into the LLM prompt
with no relevance filtering across subdirectories. This causes:

• Responses becoming increasingly long as more data sources are added
• No cross-subdirectory relevance ranking - a low-relevance document
  from one subdirectory is treated equally to a high-relevance document
  from another
• LLM receiving noisy context which reduces answer precision

Proposed Solution

Add a reranking layer after weighted_reciprocal_rank and before
returning documents in both retrieve_documents() and
aretrieve_documents().

The reranker scores all retrieved documents against the original user
query using a cross-encoder model and returns only the top N most
relevant documents regardless of which subdirectory they came from.

Implementation:

• Add src/retrievers/reranker.py using FlashRank
  (ms-marco-MiniLM-L-12-v2)
• Modify return statements in both sync and async retrieve methods
  in csv_chroma.py
• Add reranker configuration to config_default.yml

Why FlashRank:

• Runs locally - no API key required
• CPU only - no GPU needed
• Lightweight (~4MB model)
• Already compatible with existing list[Document] pipeline

Impact

• Applies automatically to both Reactome and UniProt retrievers
  since csv_chroma.py is shared
• Any future database integrations get reranking for free
• Response length directly controlled via top_n config parameter
• Zero changes to downstream pipeline - same list[Document] type
  returned throughout