Deterministically search precomputed intelligence from podcasts, TV interviews, and conference talks outside earnings calls, and from official documents published by central banks and monetary authorities (the Federal Reserve Board and regional Reserve Banks, the European Central Bank and euro-area national central banks, the Bank of England, the Bank of Japan, the Reserve Bank of India, the People's Bank of China, the Bank of Korea, the Hong Kong Monetary Authority, Taiwan's central bank, and others): speeches, testimony, policy statements, implementation notes, minutes, press-conference transcripts, projections, reports and press releases. Documents in other languages yield English paraphrases. The speakers are company executives, investors, fund managers, analysts, economists, journalists, central-bank officials, and other guests, and their claims cover listed and unlisted companies (OpenAI, DeepSeek, MiniMax), institutions (the Federal Reserve, the European Central Bank, regulators), and whole markets and industries, not only the speaker's own employer. No serving-time model interprets or expands the query.

Retrieval for structured findings is lexical, not semantic: strict full-text matching is tried first, loose any-term full-text matching only when strict matching returns no candidates, then trigram matching only when both full-text stages are empty. Natural-language questions work because person names and common question scaffolding are normalized. Prefer concise topical terms; retry synonyms or the speaker's own vocabulary before concluding there is no coverage.

Canonical search_target values:

• "search" (default): blended structured claims and broad passage cards.

• "claims": normalized decision-relevant claims only.

• "passages": broad substantive paraphrase cards only.

• "pressure_points": stored refusal and declined-to-confirm claim rows.

• "appearances": video-level catalog rows.

• "coverage": company-level structured-corpus inventory.

Compatibility aliases remain accepted for older clients: "all" maps to "search", "videos" to "appearances", and "companies" to "coverage". New callers should use the six canonical values; "search" is the primary blended target, not a legacy alias.

Result contracts:

• search/claims/passages/pressure_points return result_kind, canonical_paraphrase, speaker, primary ticker, topic labels, video/channel/publication metadata, lexical relevance, and source_url (a timestamped YouTube deep link, the episode page for a podcast, the document page for an official document, or the PDF page of a press-conference transcript). detail=true adds normalized claim and attribution fields plus institution, document_kind and event_date (null for interviews), but never raw transcript text or evidence spans.

• appearances returns video metadata, primary ticker, attribution metadata, matching claim count, URL, and relevance score. detail has no effect.

• coverage returns company-level appearance, publication-date-span, channel, claim, and low-confidence-attribution counts. It is structured coverage, not a per-person acquired-video/hour inventory; query, sort, and detail have no effect.

Filter and ordering semantics:

• company is comma-separated tickers and company names, mixed freely. A stored ticker matches the appearance's company or a section subject with that ticker; any other value is read as the name of a company, institution, or organisation a podcast section is about, matched after dropping punctuation, a leading "The", and corporate suffixes. Every ticker attached to a matched name joins the ticker filter, so "Alphabet" also reaches GOOGL rows. The result records company_tickers, company_names, company_matched, and company_unmatched (a value that matched nothing, with up to five possible matches); an unmatched value returns no rows, not an error. company_filter is the old name of company and behaves the same; passing both is an error. The appearances and coverage targets also match every value, upper-cased, against the entity references of claims (products, rivals, partners such as "CUDA"); a value found only there is listed under company_references. Alternate spellings can return different appearances or coverage rows because reference names are matched exactly ("open ai" and "OpenAI" select the same subject but not the same references); the claims rows are the same.

• person is a case-insensitive substring over the finding speaker and/or appearance speaker metadata. appearance_type applies to every target: podcast, tv_interview, conference and other for interviews; speech, testimony, minutes, statement, press_conference and hearing for official documents. country ("US") restricts every target to official documents of that country; institution ("Federal Reserve") is described with show and channel below.

• claim_family restricts structured claims. With blended search it suppresses passage cards; pairing it with search_target="passages" is an error. For appearances/coverage it requires videos with matching structured claims.

• date_from/date_to are inclusive YYYY-MM-DD bounds on the video's stored publication/upload date, which can differ from the recording or event date.

• show, channel, and institution are case-insensitive substrings of a podcast show name, a YouTube channel name, and a publishing institution; each returns only rows of its own kind, and when more than one is given rows matching any of them are returned.

• relevance orders lexical score first, then publication date; newest orders publication date first, then lexical score. Coverage has fixed inventory ordering by latest covered publication date.

• limit is one page of 1-50 rows and offset skips that many ranked rows before the page, so limit=15 with offset=15 returns rows 16-30 of the same ranking. Every result carries has_more, plus next_offset when more rows follow; call again with that offset to read the next page.

With an empty query, search and claims browse recent high-signal claims without generic passage cards; passages browses recent passage cards; pressure_points browses recent refusal rows; appearances browses the video catalog; coverage returns the company inventory. Each result set is capped at 1-50 rows.

Evidence safety: canonical_paraphrase is a normalized paraphrase, not a transcript quotation. Cite source_url (timestamped for YouTube, page-anchored for PDF transcripts) with the returned speaker, publication date, and video or document metadata. FOMC minutes describe the meeting of event_date and are published three weeks later; the Committee, not one person, is their speaker. Do not put canonical_paraphrase in quotation marks unless you independently verify the wording at the source link. The tool never returns raw transcript text.

Detail-field safety: detail adds normalized claim fields only. The extraction pipeline's internal audit flags are not published, because they describe how a row was built rather than anything checkable about the claim; judge a claim from value, assertion_status, and the timestamped source link instead. company_attribution_confidence and company_attribution_source concern only the video-to-company/ticker assignment; inferred attribution is not independent confirmation.