search_correspondence_embeds

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

**Use this to understand the REGULATORY POSITION of a filing — what the state regulator questioned, how the carrier answered, who was on the exchange. NOT to understand what the filing does or how it rates.** Wrong tool for content questions ("what does this filing change", "what's the base rate", "which forms did it introduce", "what's the indicated vs selected rate"). Reach for `search_summary_embeds` (filing content), `search_actuarial_embeds` (rate/trend/credibility numerics), or `search_filing_embeds` (paragraph-level filing body) instead. Right tool for questions about the state's dialogue with the carrier: "what did CDI push back on", "who is the reviewer on this filing", "which of my Pending Rate filings still have unresolved objections", "every filing involving actuary X". Pure vector search over `correspondence_embeds` — the per-chunk embed table populated from three attachment types in a filing's source folder: - **correspondence_attachment_*.pdf** — CDI objection letters + carrier response letters (the dialogue itself) - **supporting_document_attachment_*.pdf** — actuarial memos, exhibits, transmittal letters (highest volume of filer contact info) - **<SERFF>.pdf** (top-level cover page) — SERFF's official filing-person contact block (highest per-page email density) Every hit is a passage from one of those, ranked by cosine similarity. The excerpt IS the source text — you can quote it back to the user to replay the exchange or identify who filed the paper without a separate PDF fetch. Use `source_file` on each hit to see which attachment type the excerpt came from. **Facet filters** (all optional, all combine with AND on top of the semantic ranking): - `email` — chunks mentioning this exact address (case-insensitive) - `email_domain` — chunks mentioning any address on this domain (e.g. `insurance.ca.gov` for CDI reviewers) - `serff` — scope to one filing's thread - `filing_status` — substring match (use `Pending` for live objections, `Closed` for resolved) - `year` / `year_from` / `year_to` / `date_from` / `date_to` — filing-year and filing-date windows - `state`, `naic`, `product_type`, `filing_type`, `predecessor_prefix` — carrier + programme scope For "which filings did jane.reviewer@insurance.ca.gov touch this quarter", use `list_email_workload` instead — it aggregates and is designed for caseload views. **Cost**: one query-embedding call + one indexed Postgres lookup. Bounded, cheap, fast. **Right surface for**: - "What did CDI push back on in this filing?" — pass `serff` to scope; get the reviewer's own words. - "Show me carrier responses to territory-factor objections" — semantic search, no scope; excerpts read as regulator-carrier dialogue. - "Find filings where CDI questioned reinsurance costs" — use the semantic query alone. - "Every objection where jane.reviewer@insurance.ca.gov was on the exchange" — pass `email`, optionally combined with the semantic query. - "All filings whose objection thread involves anyone at Farmers" — pass `email_domain=farmersinsurance.com`. **Right combination with other tools**: pair with `get_filing_correspondence` to pull the full ordered thread for one filing once search surfaces a hit worth reading end-to-end. **Wrong surface for**: - Filing-content questions (rate manuals, forms, actuarial memos) — use `search_summary_embeds` or `search_filing_embeds`. - "Which filings had ANY objections at all" — for presence-only, prefer `get_filing_correspondence` with a `file_count > 0` check per SERFF. **Facet coverage** (as of 2026): most objection letters do not embed email addresses in the body — CDI reviewers sign off with a name + division, not a mailbox. So `email` / `email_domain` scopes will match a minority of chunks even for filings that had a full objection round. Semantic search is the dominant surface here; the email facets are a bonus filter, not the primary shape. Returns top-K chunks with `{serff, source_file, chunk_index, similarity, excerpt, emails, email_domains, page_date, meta}`. Default `topK=10`, max 50. Excerpt is the first 1200 chars of the matching chunk.