Phase 5c: AI Search & RAG Enhancements

[kpeatt]

Overview

Major overhaul of the RAG pipeline: query understanding, hybrid search, LLM reranking, redesigned tools, richer context assembly, conversation memory, and observability.

Tasks

• Query analysis pipeline: Build analyzeQuery() — single Gemini Flash call extracting intent, entities, temporal resolution, query rewrites. Replace get_current_date tool. Add entity pre-resolution.
• Shared embedding + hybrid search: Generate query embedding once per request. Add tsvector + GIN indexes on transcript_segments, motions, agenda_items, matters, documents. Create hybrid_search RPC with reciprocal rank fusion (RRF).
• LLM reranking: Batch top-30 candidates into a single Gemini Flash call, score 0-10. Only trigger for complex queries with >10 candidates.
• Redesigned tool set: search_discussions (replaces search_transcript_segments + search_agenda_items), search_decisions (replaces search_motions), get_person_activity (merges statements + voting), search_documents (new), get_meeting_context (new), get_timeline (new).
• Rich context assembly: Replace truncateForContext() with structured text formatting. Build full evidence bundles for synthesizer with quotes, vote breakdowns, document excerpts.
• Orchestrator improvements: Adaptive step budget by query intent. Self-correction (retry with broader query on sparse results). Parallel tool execution via Promise.all.
• Answer quality: Confidence scoring. Claim verification pass. Confidence indicator in UI.
• Conversation memory: Store state in KV (session_id → turns with evidence + entities). Pronoun resolution. Evidence reuse for follow-ups. Suggested follow-up questions.
• Observability: Per-question telemetry. Admin quality dashboard. Thumbs up/down feedback buttons.
• Search UI: Unified search bar with auto-routing. Suggested follow-ups as clickable chips. Enhanced citation hover cards.
• Fix search_agenda_items injection vulnerability: Sanitize query in PostgREST ilike filter.
• Create missing RPC functions: match_transcript_segments, match_motions, match_matters, match_agenda_items.

Current Problems

• Redundant embedding generation (same embedding generated 3x per question)
• No query rewriting — raw user text passed to vector search
• Truncation destroys context (15 items, 2000 chars cap)
• No hybrid search — vector-only misses exact matches
• No reranking — HNSW scores are approximate
• Single-shot orchestration with no self-correction
• No document search capability
• Thin synthesizer context (120-char titles, not evidence text)

Tool Redesign

New Tool	Replaces	Purpose
search_discussions	search_transcript_segments + search_agenda_items	What was discussed
search_decisions	search_motions	What was decided
get_person_activity	get_statements_by_person + get_voting_history	Everything about a person
search_documents	(new)	Staff reports, attachments
get_meeting_context	(new)	Full meeting detail
get_timeline	(new)	Chronological matter history

[kpeatt]

Partially shipped across v1.1 and v1.6

What was built:

• v1.1 Phase 8 (Unified Search & Hybrid RAG, 5 plans, 2026-02-18): Hybrid vector+text search, query embedding, RPC functions for transcript/motion/matter/agenda matching, RAG pipeline with Gemini
• v1.6 Phases 29-31 (Search Experience, 7 plans, 2026-03-01): Search backend foundation, citation UX, search controls & polish, date filtering, result cards with source citations

What remains (see new issues):

• LLM reranking
• Redesigned tool set (search_discussions, search_decisions, etc.)
• Conversation memory (KV session state, follow-ups)
• Observability (per-question telemetry, admin dashboard, feedback buttons)
• Query analysis pipeline improvements
• Confidence scoring and claim verification

Closing this umbrella issue. Remaining items tracked as separate focused issues.