search_filing_embeds

ai.serff/ca-rate-filings · Verify this server

Pure vector search over per-chunk full-document embeddings (`filing_embeds`, ~12.4M rows across ~65K filings — each filing sliced into ~190 paragraph-sized chunks). The most granular semantic surface in the corpus. **Cost**: one query-embedding call + one indexed Postgres lookup. No LLM planning, no LLM composition. **Right surface for**: - "Find the exact passage discussing X" — granular text-search where you need the paragraph not just the filing. - "Find filings whose body text mentions X" when the summary-level surface (`search_summary_embeds`) might miss a topic buried in a long PDF. - **"Drill into this specific filing semantically"** — pass `serff` to restrict the cosine search to a single filing. Without scoping, commodity-vocabulary chunks from other filings can out-rank your target filing; scoping eliminates that. **Wrong surface for**: - Filing-level questions where multiple hits per filing are noise — use `search_summary_embeds` (one match per filing). - Concrete-filter questions like "Filings from carrier NAIC 12345 in 2024" — use `search_filings`. `aggregate: true` (default) collapses to top-K *filings* by best-chunk similarity (one row per filing, the best matching paragraph as excerpt). `aggregate: false` returns top-K raw chunks (may include several from the same filing) — use when the user asked to see the actual paragraphs. When `serff` is set, aggregate is forced to false (every hit is the same filing already). Returns top-K hits, each with `{serff, chunk_index, similarity, excerpt, meta}`. Default `topK=10`, max 50. Excerpt is the first 800 chars of the matching chunk.