Knowledge Base

There is nothing to upload. The Chatfuel knowledge base is a structured record — no file store, no crawler, no chunker, no index, no embeddings, no retrieval endpoint and no similarity score anywhere in the API — so this module is an editor over that record, and anything in it that looks like retrieval quality is local, lexical and says so.

Eight sources in a rail beside the selected source's page: what the business is, what it answers, and what it could not answer yet.

Installed as

Routes and views

This module has one route and no view segment. /knowledge-base is the whole surface, and ?source= picks which of the eight the rail has open — the default, overview, is omitted from the URL.

Services and Team are one entity seen from two modules. When bookings is installed it owns editing them, so they render read-only here with a link to /bookings/services and /bookings/staff; a scaffold that took this module without bookings gets a working editor instead of a link to nowhere.

The rest of the parameters:

A ?tab=business|faqs|catalog from the three-tab page this replaced is still read, mapped onto source, and dropped on the next write. An unrecognised value anywhere renders the default rather than throwing.

The model

One record, read as eight sources:

Everything in that record is fed to the assistant as text on every conversation, so the character budget is the constraint everything else competes for. The server reports two numbers — usage.total and usage.catalog — and no ceiling. There is no limit field in the schema, which is why nothing in this module draws a gauge: the meter shows composition, which source is spending what, and "full" arrives as FuelyKnowledgeBaseLimitReached on a failed write rather than as a threshold anyone guessed.

Writes are granular in one place and wholesale in two others. The business profile has one mutation per field and no bulk update. Opening hours are a full replace — all seven days go every time. And the FAQ list is replaced whole:

One field is misnamed on the server and it matters: additionalInstructions is not deprecated and is not a prompt. The behaviour prompt moved to the per-scope automation settings; what is left under that name is the free-text half of the business, which is why this module calls the source "About the business" rather than "AI instructions".

Configuration

Nothing here is specific to this module. Environment variables has what each one is and which side reads it.

Permissions

A role lookup that fails is treated as permitted rather than denied: the API refuses the write anyway, and locking someone out of a page they are allowed to use is the worse failure.

Limits

There is no ingestion API. An import is a local parse — a CSV, TSV, TXT or Markdown file read in the browser, or pasted text — that produces rows a person reviews before they are written through the ordinary mutations. Export is the mirror: CSV for a spreadsheet, JSON for everything else, and a read, so someone who cannot edit can still take their data.

Reading a page from the customer's website is deliberately not offered. A browser cannot read a cross-origin page, so the only way to offer it is a server endpoint that fetches an address a person typed from inside the deployment's network — which is the shape of an SSRF bug. What a URL buys over pasting is one copy and paste. Weigh that against the surface before adding it back.

One budget, no ceiling. Every source that spends characters competes with every other, and the only way to learn you are full is a failed write. Plan the UI around composition, not around a percentage. The catalog has a second, separate ceiling with an error of its own: GoodsItemsTooMuchForBot is too many items, FuelyKnowledgeBaseLimitReached is too many characters, and any goods mutation can answer with either.

Nothing on this page is live. The schema has exactly one Fuely subscription and it belongs to the automations domain, so freshness here is: merge each setter's own response, refetch on reconnect, and a Refresh control.

The persona fields are deprecated and always come back empty. agentName, chatLanguage, greeting, messageLength, emojiPolicy and the rest are still in the schema, so a client can select them — and then renders blank inputs on every real bot, which reads as a bug in your code. They moved to the per-scope automation settings, along with the rest of how the AI behaves.

Gaps has two signals and no more. unhandledSwitchToHuman on a contact, and an assignee that is a person rather than the AI. There is no resolution flag, no CSAT, no "the AI said it did not know" marker and no server-side grouping — so the clustering is a word overlap, is described to the reader as a word overlap, and prints counts of real conversations rather than a confidence score. It also cannot see a hand-off somebody already handled (unhandledSwitchToHuman flips false the moment an operator opens the chat), a customer who gave up and left without a hand-off, or whether an answer the assistant did give was any good.

The module's own ceilings: