Reader job: discover related disclosure language without losing provenance
Start with a specific question and a constrained issuer or filing universe. Preserve the candidate’s source identity, then retrieve its section when the application needs to quote, summarize, or compare it. That makes semantic search a route to evidence rather than an opaque answer generator.Search a constrained question
q is required. Narrow the candidate set with ticker, symbol, or cik, then add form and filing_year when the research warrants it. mode accepts keyword, semantic, or hybrid; hybrid is the default. limit bounds one response.
A result can include accession, section key, source URL, ticker, highlighted snippet, and char_start/char_end offsets into normalized section markdown. Default responses can include score, retrieval mode, and a degraded-state signal. Retain these fields with the query filters, requestId, and any provenance or freshness metadata.
Verify the candidate section
_citation_degraded is returned, open the source filing instead of constructing a quote from a partial excerpt.
Capability choices
The JavaScript SDK provides
client.semanticSearch({ q, ticker, form, filing_year, mode, limit }). Hosted MCP offers normalized sections.search, but not an equivalent semantic-search tool; use REST or an SDK for semantic retrieval, then hand source identity into the agent context.
Retrieval limits
- Similarity and hybrid scores are not materiality scores, completeness measures, or recommendations.
- Results reflect indexed, normalized section content and current route availability, not a promise of EDGAR-wide prose coverage or coverage for related passages.
view=agentkeeps citation fields but omitsscoreandretrievalMode; use the default response when evaluating retrieval behavior.- A
503or reported degraded state is an unavailable or partial retrieval surface, not a negative result.

