RAG prototype: Retrieval-Augmented Generation for Foliofox

[unav4ila8le]

Title: Retrieval-Augmented Generation (RAG) Prototype for Foliofox

Summary:

• Proposal: Implement a RAG-based assistant to provide grounded, personalized financial advice by combining vector memory of user sessions and curated personal-finance knowledge with our existing model wrapper.

Benefits:

• Grounded, citation-backed answers (reduces hallucinations).
• Personalized advice using user-specific memory (preferences, past sessions, portfolio snapshots).
• Faster and cheaper responses by retrieving compact context instead of re-sending large raw data.
• Auditability: each recommendation can cite the source(s) used.

Proposed minimal plan (MVP):

1. Phase A — Prototype (1-2 weeks):
   • Ingest: 30 session summaries + 30 curated finance articles.
   • Store embeddings in a small vector index (per-user namespace + global index).
   • Build retriever that returns top-3 items and wire into RAG prompt template.
   • Test with 5 example queries and compare answers vs current approach.
2. Phase B — Personalization (2-4 weeks):
   • Automate session summarization and per-user indexing.
   • Add UI to surface 'used sources' and allow user feedback (thumbs up/down).
3. Phase C — Production & Policies (ongoing):
   • Add pruning, retention policy, encryption/opt-in controls, and CI smoke-tests.

Success metrics:

• Reduced hallucination rate (manual spot-checks).
• Increased user trust (feedback), lower model cost per query, faster response time.

Security & privacy:

• Store per-user vectors in isolated namespaces; offer opt-in and delete-on-request.
• Avoid storing raw sensitive credentials or PII; store summaries instead.

Request:

• Approve prototype and assign to me for an initial implementation PR, or tell me who should own it.