agenthub · document.fetch
Documentation Fetch
Ground the agent in the business's own Document Space, so it answers from documented facts rather than improvising.
Why the node earns its place
Why a node and not a tool: a tool is optional. The model decides whether to call it, and when it skips, the agent invents an answer. A node is deterministic — the documentation is in context before the model generates a token, so it cannot answer ungrounded. For a lead-qualification agent, where a fabricated price is a real business cost, that guarantee is the whole point.
It sits exactly where Knowledge Fetch sits and mirrors its ports, so the two are interchangeable. The difference is the source: this reads the authored markdown directly, so an edit lands on the very next turn with no re-index step.
How it works at run time
- 1Read the org from the run and list the documents bound to this agent.
- 2In
allmode inject every bound document; insearchmode keyword-filter by the latest user message first — only worth it once the corpus outgrows the context budget. - 3Respect
max_documentsand themax_charsbudget. - 4With
stricton, append an explicit instruction to answer only from the documentation and to say plainly when it does not cover something.
What you wire
Inputs
- messages *array
- system_promptstring
Outputs
- messagesarray
- textstring
- documentsarray
Configuration
| Key | Type | Default & options | What it does |
|---|---|---|---|
| system_prompt | string | — | Agent persona / instructions. Prepended to the documentation block inside the same system message. |
| mode | string | default "all" all · search | all: inject every bound document (right for a normal doc space). search: keyword-filter by the latest user message first — use only when the corpus outgrows the context budget. |
| max_documents | number | default 50 | Cap on documents injected. |
| max_chars | number | default 24000 | Total character budget for the documentation block. |
| strict | boolean | default true | Append an explicit instruction to answer only from the documentation and to say so when it doesn't cover something. |
What usually goes wrong
Watch for this
Failures are swallowed on purpose — retrieval must never block the chat path. On a missing binding or an error the agent falls through to persona-only and keeps talking, ungrounded. An agent that has quietly started improvising is the symptom; check the document bindings rather than the prompt.
Business accounts only. An individual account has no business documentation, so the node makes the agent say so rather than degrading into a generic chatbot.
Behaviour & provenance