Should I use Knowledge Search or a Resource?#
Use Knowledge Search when an agent needs to read documents it receives in a run; build a Resource when many agents and callers need the same reviewed data, versioned and kept fresh.
| Knowledge Search node | Resource Foundry | |
|---|---|---|
| Best for | Answering from documents passed into one run | A reusable, reviewed dataset many agents query |
| Input | Text, extracted PDFs and Word files, fetched pages | Manual text, JSON rows, or a public URL or site |
| Output | The most relevant passages, ranked with scores | Exact records with only the fields you allow |
| Review | None needed: it reads what you hand it | You approve every pack version before it runs |
| Where | A node on the canvas | Resources in your workspace |
How does the Knowledge Search node work?#
Knowledge Search splits your documents into passages, ranks them against the question by keyword relevance, and returns the best matches with their scores. It runs inside the run itself, with no vector database to set up and no outside service to call. Feed its output to an LLM (Claude) step and the answer quotes your material.
| Setting | Default | Range |
|---|---|---|
| Query | The question to answer | Any text, or a value from an earlier step |
| Documents | Whatever the previous step hands over | Up to 25 documents |
| Top results | 3 passages | 1 to 20 |
| Passage size | 150 words | Up to 2,000 words |
Build an agent that answers from your documents
Step 2: Bring in the documents
Add Extract PDF Text, Extract DOCX Text or Fetch a URL to turn your files and pages into text.Step 3: Add Knowledge Search
Connect the extracted text to a Knowledge Search node and map the user's question into Query.Step 4: Answer with Claude
Pass the top passages to an LLM (Claude) step with an instruction to answer only from them, then test the agent for free.
How do I build a Resource in Resource Foundry?#
Resource Foundry turns material you choose into a reviewed Resource Pack that does one narrow job well. Open Resources and press new, or go straight to Frame one resource-backed job. Pick a Starter brief (Niche Data Refinery, Agent Readiness or Expert Archive), then fill in the brief.
| Field | What to write |
|---|---|
| Resource name | A clear name buyers recognize |
| Narrow job | The one question this resource answers |
| Buyer or buying-agent intent | Who asks it and why |
| Execution access | Free, Paid or Private |
| Discovery access | Unlisted or Public |
| Typed record fields | Fields and types, such as name:string, score:number |
| Filter fields and Return fields | What callers filter by, and the only fields results show |
| Taxonomy tags | Labels that group your records |
Every resource then moves through seven stages: Brief, Sources, Records, Job, Test, Publish and Trust & Earnings. Add sources as manual text, JSON rows or a public URL. Review the proposed schema, records, evidence, unknowns and conflicts, then approve. Approval freezes that pack version so it never changes under a live agent.
How do I test, publish and refresh a Resource?#
Test the exact approved pack before it goes anywhere. The test returns the pack's version, freshness, evidence, unknowns, conflicts and whether the output matched its schema. Publishing runs through the same Test-to-Live path as every agent, so you always know which version is live.
Agents read a published pack with the Resource Query node. It queries one exact approved version, accepts only your declared filter fields, returns only your declared return fields, and sends back up to 100 records per call. When your source changes, a refresh builds a new candidate and shows you the differences while the live release keeps serving. Approve the candidate and publish it when you are ready.