Data Model Overview

All persistent data lives in Supabase (Postgres 17) with pgvector for embeddings. Row-Level Security is enabled on every table.

Entity Relationship Diagram

Table	Rows	Purpose
`transcript_chunks`	11,546	Chunked transcripts with `halfvec(1536)` embeddings + HNSW index
`episode_metadata`	345	Podcast episode info (title, speakers, audio_url, published_at)
`persons`	24	Speaker profiles (name, bio, wiki, trust badge, domain scores)
`users`	—	App users (linked to Supabase Auth)
`threads`	—	Search sessions (max 5 active per user)
`messages`	—	Chat messages within threads
`thread_beliefs`	—	Saved beliefs per thread (unique per chunk_id + thread_id)
`community_cards`	—	Published community content
`community_card_votes`	—	Vote tracking
`watchlist_items`	—	User watchlist
`alerts` / `notifications`	—	Alert system
`beta_invites`	—	Beta access codes
`speaker_claims`	—	User→speaker verification claims

Function	Purpose
`search_transcripts(embedding, count, speaker?, episode?)`	Vector similarity search with HNSW
`get_chunks_by_speaker(speaker, limit)`	Transcript chunks for a speaker
`get_episode_by_id(id)`	Episode metadata lookup
`get_speaker_aggregations()`	Speaker nodes for graph visualization
`get_speaker_similarity(min_sim)`	Speaker-to-speaker similarity edges
`get_speaker_coappearances(min_shared)`	Co-occurrence links
`get_timeline_data(speaker)`	Speaker mentions over time
`get_speaker_topics(speaker)`	Topics a speaker discusses
`get_related_speakers(speaker)`	Similar speakers

CREATE INDEX ON transcript_chunks
  USING hnsw (embedding halfvec_cosine_ops)
  WITH (m = 16, ef_construction = 64);

Type: HNSW (Hierarchical Navigable Small World)
Operator: halfvec_cosine_ops (cosine similarity on half-precision vectors)
Parameters: m=16 (max connections per layer), ef_construction=64 (build quality)
Dimensions: 1,536 (halfvec(1536))

23 migrations track the full schema evolution from initial functions through community features:

Range	Feature
000-006	Core: functions, users, threads, messages, beta invites, speaker claims
007-011	Transcripts: tables, search functions, foreign keys, RPC hardening
012-013	Community: cards and votes
014-018	Threads: thread model, beliefs, watchlist, alerts/notifications
019-023	Polish: shared threads, speaker aggregations, deduplication, public access