How it works
How it works
A pipeline run covers one business and attempts every source. It writes the safe results first and runs the LLM step last, so a failure late in the run never loses earlier work.
- Step 1
Create a business
Enter a name and a domain. The tool gives the business a bizId and a slug, and adds both to the business index (
data/index.json). - Step 2
Start a pipeline run
The run takes a lock on the business, then attempts each source in turn: the website, BBB, social discovery, Google Places, GDELT, Serper, D&B and NewsAPI. Each one records a source status.
- Step 3
Capture and extract claims
Each connector returns claims. A claim names its field path, value, source, capture method, time and confidence. A blocked page stops there; a person can capture it by hand later.
- Step 4
Merge and resolve
Fresh claims are merged with the ones already on disk, then resolved: your curation decision first, then the highest confidence. Disagreements at the top stay open as conflicts.
- Step 5
Write durable data
data.json,claims.jsonandmeta.jsonare written before the LLM step runs. A source that failed this run keeps its earlier data untouched. - Step 6
LLM curation pass
With an Anthropic or OpenAI key, the tool summarises the resolved values into
profile.mdand flags conflicts for you. Without a key, this step reportsnot_configured. - Step 7
Review and decide
In the Curation tab you accept a claim, edit a field or reject a field. Each decision is saved in
curation.jsonand beats every later run.
Safe to run again
Re-running a business is expected. A run that finds nothing new produces a byte-identical data.json.
A source that is blocked or fails this run never overwrites what it captured before; its status in meta.json says missing_this_run or names the failure.
The tool never writes curation.json itself. Only you do, from the Curation tab.
Research the next business in minutes
The tool runs locally from a clone of the repository. Access to the repository is by request.