Research

Request question research

Enqueue or reuse bounded public-evidence research. Checks saved originals first. Explicit URL and date constraints can select stored originals even without matching words in their body. No scraping in enqueue. Never approves Lore or attaches a profile. Newest-available retrieval first checks stored dated public evidence, including own social captions and responses already published in Latest. Stored Spotify/Deezer release catalogs and InProcess moments share durable snapshots with the public Latest endpoint. Provider refreshes run only in queued latest_refresh work; question/status/original reads never refresh these providers. A bounded agent explores the saved public snapshot with paginated browsing, model-chosen searches and exact source-window reads before selecting unchanged evidence; exact supporting quotations and factual claims are checked during answer drafting. If stored evidence is insufficient, the model chooses a next lookup from server-validated web, page or connected social options based on the missing information. There is no fixed social-platform priority. Explicit URL/platform restrictions, account identity and existing provider budgets still constrain execution; the chosen plan is persisted and revalidated before subsequent job slices. Web pages retain explicit publication metadata when unambiguous; unknown dates remain unknown, never replaced by fetch time. Every new job reserves one slot in the rolling global saved-question admission cap (100 per 24 hours), with no per-artist saved-chat cap. This is an operational request limit, not a dollar spending guarantee. Direct or saved-source answers consume no outside-research slot. Only when insufficient saved evidence leads to a validated external lookup does the server reserve outside allowance (5 per artist and 100 globally per rolling 24 hours). That reservation is atomic, lease-fenced, and idempotent for the same job; all existing per-job provider/model/lifetime limits still apply. Exhausted outside allowance ends the job unresolved without starting a provider, with outsideResearchReason=quota plus quotaScope, retryAfterSeconds and resetAt. Exhausted global saved admission returns saved_evidence_quota (429), quotaScope=saved_global, and the same reset fields; Retry-After matches retryAfterSeconds. Reset timing comes from the actual rolling-window boundary and is an estimate assuming no newer reservations consume capacity. Old jobs retain their original quota accounting until their rolling window expires. Single-active-job research_busy remains distinct. Completed matching jobs can be reused independently of these admission limits, subject to current eligibility and cache lifetime. Cache identity preserves topic and all source/date/evidence constraints; reuse is not restricted to an arbitrary last-ten-jobs window. Saved research reads all eligible originals within the existing corpus guard, independently of the Latest UI card limits. Published answers retain publication eligibility and private answers remain excluded. Partial saved support is revalidated and combined with newly collected originals before final assessment. Saved selection returns at most 12,000 original characters; combined outside assessment can return at most six windows totaling 24,000 characters. Exploration uses a current public-research model through AI Gateway, with per-step reservation and usage tracking; external provider budgets are unchanged. Service callers may opt into preferDirectEvidence: true to offer bounded saved candidates before model-driven evidence exploration. This option is not part of the persisted neutral request. Selection candidates means evidence is ready for a checked answer attempt, not that it answers the question. At most six exact windows and 12,000 characters are selected. Latest reporting overviews select the six newest distinct canonical posts first, preferring written evidence over a duplicate transcript, then interleave providers only within that cohort. For overview follow-ups with excluded source URLs and no target URL, known In Process posts from the same connected account and stable collection URL contribute their newest eligible representative before the six-candidate limit. Other social posts remain separate activities. Collection names and question keywords do not determine grouping. This diversifies candidate evidence without treating an entire collection as excluded or answered; focused requests and overviews without exclusions keep individual-post selection. Explicit source, platform, evidence-kind, date and URL exclusions still apply. Semantic project exclusions remain an answer-layer obligation. An empty selection enters ordinary research. Candidate jobs reserve the same quota once and preserve normal claim and expiry checks. Legacy callers never reuse a candidate-only job. A client must register and verify its answer, reread current evidence, and continue the same job if the candidate answer is insufficient or fails verification.

curl --request POST \
  --url 'https://musicnerd-api.vercel.app/api/artist/{id}/research/questions' \
  --header 'Content-Type: application/json' \
  --data '{
  "topic": "string",
  "evidenceNeed": "reporting"
}'