Knowledge base
For an agent, a knowledge base is the set of material a workspace declares authoritative — the things an answer is allowed to be built from. That is a narrower idea than a document library. A page nobody has approved is not knowledge, and a folder of eighty overlapping drafts makes an agent less accurate rather than more.
Six properties that decide whether it works#
- Authority. Somebody owns each source and says it is current. Unowned material ages into a wrong answer.
- Granularity. Retrievable units are about one thing. A forty-page handbook is one document and forty topics; retrieved whole, it drowns the question.
- Non-contradiction. Two sources that disagree turn every answer into a coin toss the reader cannot see.
- Question-shaped language. Sources written in the customer's words are found by the customer's questions.
- Coverage that is known. Knowing what the base does not cover is worth as much as the contents, because it is the boundary the agent must decline at.
- A correction path. When an answer is wrong, changing the source must change the next answer. If corrections live in replies, nothing has been learned.
Size appears on no line of that list. A tight base of forty accurate, single-topic sources outperforms four hundred documents of mixed vintage, and does so on every retrieval mechanism, because the second base contains its own contradictions.
Does Connect use a knowledge base?#
Used. Knowledge is a first-class part of the product, not an add-on. knowledge_bank.py holds the sources a workspace supplies; knowledge.py holds Facts, each one grounded statement drawn from a source or from a conversation. Both are workspace-scoped rows under the same isolation as everything else, and both are visible and editable on the Knowledge screen.
The Facts layer is the part worth understanding. A document is a container; a Fact is a claim. Extracting claims from documents means an answer can cite the specific statement it relied on rather than the file it appeared in, a wrong claim can be corrected without editing the source it came from, and the retrieval budget can be spent in units of meaning — on the voice path, at most four facts and 2,000 characters — rather than in arbitrary slices of text.
Files a workspace uploads are handled by the shared file service rather than by a second store: PNG, JPEG and WebP images, PDF (its text layer where there is one, otherwise the model reads the page), DOCX, XLSX, CSV and TSV, PPTX, TXT, MD and JSON. Declarations that would pull in external entities, and archives that expand disproportionately, are refused at the boundary. Files and data in Connect covers that path.
Keeping one honest#
Start from the questions you already answer, not from the documents you already have.
Result The base matches demand from day one instead of matching whatever happened to be written down.
Give every source an owner and a review date.
Result Staleness becomes a visible queue rather than an incident.
Treat every declined question as a finding.
Result A gap the agent reported is a gap you can close; a gap it answered over is one you will hear about from a customer.
Delete rather than accumulate.
Result Superseded material does not become harmless by being old — retrieval has no way to know it lost an argument.
What it is not#
- A CRM
- Knowledge is what is true about the business. What is true about a person lives in relationships and memory.
- Training data
- Nothing here is used to train a model. Sources are read at answer time and can be deleted with immediate effect.
- A search index for people
- It is built to be selected from by an agent under a tight budget. That is a different optimisation from a search box a person scrolls.
- A transcript archive
- Conversations are records in their own right. Knowledge is the material answers are allowed to rest on.
Questions#
How much should we put in before switching a channel on?
Enough to cover the questions you receive most often, and no filler. A base that answers the top handful of recurring questions accurately and declines the rest is immediately useful; one padded to look complete produces confident answers in the areas you never checked.
What happens to an answer when we edit a source?
The next answer uses the edited material — sources are read at answer time, not compiled into a model. There is nothing to re-index and no delay to wait out, which is a direct consequence of grounding over records rather than over an embedding index.
Can a customer's email become knowledge?
A Fact can be extracted from a conversation, which is how something learned in a thread becomes reusable. It is a Fact with its own provenance, editable and removable like any other — not a silent edit to a source document.