> ## Documentation Index
> Fetch the complete documentation index at: https://docs.noisemaker.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Research pipeline

> Build traceable market and ICP artifacts for prompt and recommendation work.

Research turns workspace configuration and external evidence into immutable artifacts. A completed research run is required before current prompt regeneration.

## Requirements

* Products and services must be filled in.
* The operator must provide OpenRouter and Exa credentials.
* The workspace needs enough competitor and ICP configuration for the requested stages.
* The credit operation must reserve successfully.

An on-demand research run is limited to one start per workspace per hour.

Research also uses the shared workspace job slot. A running measurement, recommendation-generation job, or SEO refresh blocks the start before the research worker launches.

## Five stages

<Steps>
  <Step title="Product and service map">
    Builds a structured map from the immutable workspace settings staged for the run.
  </Step>

  <Step title="Competitor discovery and profiles">
    Combines configured competitors with Exa evidence derived from Search Console queries.
  </Step>

  <Step title="SERP topic map">
    Uses product queries and Search Console queries to collect and group Exa search evidence.
  </Step>

  <Step title="Market summary">
    Summarizes product, competitor, analytics, and authority evidence. Missing GA4 evidence is recorded as a risk rather than filled with a guess.
  </Step>

  <Step title="ICP query patterns">
    Produces query patterns for each ICP from completed upstream artifacts and Search Console evidence.
  </Step>
</Steps>

The artifact registry includes <code>product\_service\_map</code>, <code>competitor\_profile</code>, <code>serp\_topics</code>, <code>market\_summary</code>, <code>icp\_query\_patterns</code>, and <code>site\_audit</code>. The five-stage worker produces the market and ICP artifacts described above. Technical site audits have their own pipeline.

## Artifact contract

Each artifact includes:

* A strict kind and schema version
* The research run and credit operation identifiers
* UTC observation times
* Source envelopes with signal, URL, or upstream artifact references
* Canonical content and source digests
* A SHA-256 artifact digest

Provider attempts are also persisted. Once an attempt has responded, it is immutable. An ambiguous outcome is not automatically resent because that could duplicate provider cost.

## Completion and eligibility

Only artifacts from a completed research run can appear in the latest eligible view. Artifacts written before a failed run are retained for diagnosis but cannot feed prompt regeneration.

Research becomes stale seven days after completion. The UI can still show the old trace, but prompt regeneration refuses it.

<Warning>
  A completed artifact proves what the pipeline observed and derived. It does not prove that every external source was available or current. Read source timestamps and recorded risks.
</Warning>

## Trace view

Use a run's trace to follow:

```text theme={null}
workspace settings and signals
  -> provider attempts
  -> stage artifacts
  -> frozen selection
  -> generated prompts
```

Matching digests establish lineage. They do not imply that later prose was copied verbatim from a source.