How research works

POST /v1/research is built to feel like one call but deliver the quality of a hand-curated brief. You send a question, you get back a synthesised answer with citations — typically in 30 to 90 seconds.

Under the hood, Brime runs an iterative agent that plans the right sub-queries, gathers fresh evidence from across the open web, and writes a tight answer with every claim sourced. You never need to choose models, tune knobs, or wire up a multi-step pipeline yourself.

What you get back

Every research response carries the same shape:

• answer — a synthesised, sourced answer to your question.
• sources — the ranked evidence the answer is built on, with title, URL, snippet, score, and published_date where available.
• steps — a compact trace of which sub-queries fired in each round, useful for debugging and observability.
• request_id / credits_used / latency_ms — the metadata your pipeline needs for tracing, billing, and SLOs.

Quality guarantees

• Multi-round refinement. When the first pass doesn't have enough signal to answer confidently, Brime keeps researching until it does — up to a tuned round budget. You don't pay extra for the extra rounds.
• Citations on every claim. Every sentence in the answer ties back to one or more sources. Hallucination is minimised by design: the synthesiser only sees evidence we actually retrieved.
• Freshness aware. Recency is part of the planner's signal. Ask about something that happened today and the agent favours today's sources over older ones, automatically.
• Citation-clean answers. No mid-paragraph footnote markers cluttering the prose — references live in the sources array, keyed by index.

Latency you can plan around

Typical wall-clock: 30–90 seconds end-to-end. Two factors move the needle:

• Repeating the same (or near-same) question hits a warm cache — identical follow-up calls complete in well under a second.
• Open-ended questions with vague scope take longer than tight, well- scoped questions. "Compare X vs Y for use case Z" is faster than "what should I know about X".

If you need to surface progress to a human, streaming is built in.

Streaming, when you need it

Add Accept: text/event-stream (or stream: true in the body, or use the SDK's research_stream method) and Brime emits a live event for every step. You can render a "Researching: …" status, paint sources as they're discovered, and reveal the answer the moment it's ready.

The event vocabulary is small and stable:

If the connection drops, reconnect with the standard SSE Last-Event-ID: <n> header — Brime replays anything you missed, then continues live. Drop-tolerant by default.

For the full event schema (including error and timeout shapes), see Status & streaming.

Costs and limits

• Flat 5 credits per call, regardless of how many rounds the agent decides to run. No per-round surprises.
• Polling status or streaming events is free — no double-billing for the long-running view.
• The single bound on call duration is a 10-minute soft alarm. Practically, every successful call finishes well within that.
• Errors are never billed. A 401 / 429 / 5xx response carries credits_used: 0.

When research is the right tool

If you find yourself orchestrating search → extract → LLM-synthesis by hand, that's the loop research replaces.

See also

• Research API reference — request and response schemas with the live playground.
• Best practices: research — query shaping, filtering, and edge cases.
• Credit model — why 5 credits is the flat price.