agenthub · knowledge.fetch
Knowledge Fetch
Semantic search over the org's pre-ingested knowledge, injected as grounding context before the model answers.
Why the node earns its place
The vector-search counterpart to Document Fetch. Where that reads authored markdown whole, this searches an ingested chunk index and injects only the passages nearest the question — which is what you want once the corpus is larger than any context budget.
How it works at run time
- 1Take the latest user message as the query.
- 2Embed it, and ask for the nearest
top_kchunks. - 3If the embedding call fails, fall back to lexical search — a weak match beats no answer.
- 4Inject as one system message, or one per chunk, per
inject_as.
What you wire
Inputs
- messages *array
- system_promptstring
Outputs
- messagesarray
- textstring
- chunksarray
Configuration
| Key | Type | Default & options | What it does |
|---|---|---|---|
| system_prompt | string | — | Agent persona / instructions. Prepended to the retrieved knowledge block inside the same system message. |
| top_k | number | default 6 | Max chunks to retrieve |
| inject_as | string | default "system" system · prepend | system: collapse chunks into one system message before the incoming messages. prepend: insert one system message per chunk before the incoming messages. |
What usually goes wrong
Watch for this
Like Document Fetch, failures fall through to the persona-only path rather than blocking the reply. Chunks have to be ingested first — the crawl, filter, chunk, embed and store chain is what fills the index this node reads.
Behaviour & provenance