Skip to content

AI/ADES source intelligence ​

New corporate/SEC, financial/action and scientific research records are processed through a shared durable queue. ADES resolves entity and market-graph candidates; the configured Qwen model interprets the source and a separate model pass reviews grounding. Failed model/ADES calls remain retryable. Raw collection does not mean digestion is complete.

Find source coverage ​

GET /v1/intel/digests lists public source keys and coverage status. Filters: kind, id, limit (1–100, default 25) and opaque numeric cursor. Supported kinds: company_profile, corporate_observation, financial_fact, financial_document, corporate_action, research_paper, research_enrichment. Collections return data, nextCursor and schemaVersion. The same API-key and quota rules apply. This includes document and bibliographic-observation digests beyond the primary source endpoints.

Read a digest ​

GET /v1/intel/digests/:key uses the same bearer API key, email-verification and quota rules as the other intelligence endpoints. URL-encode intelligence.sourceKey from a source response. Unknown and internally restricted keys return 404.

The response contains source identity/revision and data:

FieldMeaning
sourceKeyImmutable source identity and revision key
statuspending, processing, retry or complete
completedChunks, totalChunksCoverage of all source-text groups; zero total means awaiting scheduling
resultsUp to 20 completed group analyses; partial results do not imply complete source coverage
moreResultsAdditional retained group results exist; the initial public response is bounded

Each result includes summary, assessment, entities, impacts, limitations, processing timestamp, model, ADES/artifact versions, source keys, evidence references and grounding-review receipt. Original source text is withheld; follow evidence sourceUrl links. Financial observations may share a group analysis; sourceKeys identifies its scope, while returned entities/impacts are filtered to those citing the requested record. The group summary may describe neighboring observations and is not an individual fact's conclusion.

assessment is potential_impacts, insufficient_evidence or no_material_impact. The latter two are completed, valid analyses. API consumers must not interpret missing impacts as a failed job or invent a substitute score.

Interpretation ​

Entities distinguish mentioned from potentially_affected. canonicalRef comes from actual ADES output; unresolved mentions use null. Impacts are explicitly modeled_potential, with direction, horizon, conditional mechanism, assumptions, counterfactors and evidence IDs. Model confidence is uncalibrated and capped at 0.8; it is not a forecast probability. Catalog associations are not evidence that an asset price will move. Tickers must come from retained company metadata or ADES candidates.

Old publications and source revisions remain permanent. Newly generated analysis is additive. Existing records are backfilled asynchronously; all accepted normalized records are tracked, including records that never become posts. Financial snapshots are immutable source containers; their normalized financial observations are covered individually through grouped receipts. Provider raw fields that were not accepted as normalized observations are not advertised as analyzed financial facts.

The initial public response caps long-document result display at 20 groups while retaining complete coverage counts and all internal results. This does not truncate input or stop processing later groups. New source records receive priority by source timestamp; older backfill remains queued under bounded throughput. Internal-only announcements are processed internally but do not expose source bodies or digests through the public API.

The Neural Ledger API