> ## Documentation Index
> Fetch the complete documentation index at: https://docs.vorel.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Knowledge
> What the agent answers from: documents, imports, your catalog, collections, and the gaps the platform finds for you.
The agent answers from your knowledge, not from general world knowledge. If the answer is not in
here, it says so and offers a human.
## Getting knowledge in
Add an entry with a question and an answer. Best for the twenty things customers always ask.
Price lists, policies, service descriptions, menus. Content is extracted and indexed.
Point at a page on your site and pull the content in.
Load many entries at once from a structured file.
Everything lands in one library at **Knowledge**, where entries can be edited in place, grouped and
retired.
## Collections
Entries can be grouped into collections, which is how you keep one workspace's knowledge organised
when it covers several branches, brands or service lines.
## Your catalog is separate
Knowledge answers questions. Your **catalog** (offerings, properties, services, menu items) is
structured data the agent searches and books against. A price lives in the catalog; the policy about
how prices work lives in knowledge.
Catalog attributes are seeded by your [vertical pack](/verticals/generic): real estate ships
bedrooms, bathrooms and size; salon ships duration and category.
A catalog item that has not finished indexing cannot be found by a semantic search, only by exact
filters. If the agent says it cannot find something you know exists, check that the item finished
importing.
## How retrieval works
A customer question is matched against your knowledge semantically rather than by keyword, so
"do you take walk-ins" finds an entry titled "appointment policy". Each match carries a similarity
score, and the agent will not answer from a match that scores too low. It escalates instead.
The [explainability panel](/agent/explainability) shows exactly which entries were retrieved for a
given answer and how closely each matched.
## Knowledge gaps
The platform mines your own failures each week and tells you what you have no answer for.
It looks at questions that led to hallucination flags, escalations, thumbs-down ratings, and
unanswered turns, clusters them by topic, and scores each cluster by how often it comes up, how
recently, and how badly it went. What you get is a ranked list of missing answers rather than a
transcript dump.
Each finding is one of two kinds:
| Kind | Meaning |
| ------------------ | ----------------------------------------------------------------------------------------------- |
| **Content gap** | You have no answer for this. Write one. |
| **Embeddings gap** | You do have documents on this, but they never finished indexing, so the agent cannot find them. |
The distinction matters: a cluster of failures does not prove the answer is missing. Sometimes the
answer exists and is unreachable, and writing a second copy of it would not help.
You can ask the platform to draft an answer for a content gap, review it, and publish it into your
knowledge base. Drafting is on demand rather than automatic, so nothing is written in your voice
without you asking for it.
A weekly count of new gaps appears in your digest.
## Connectors
Knowledge can be pulled from external sources you connect rather than uploaded by hand. Available
sources are listed under **Knowledge, then Sources**, and each one is connected with its own
authorisation.
## Keeping it current
* **Retire rather than delete** when an answer stops being true, so you keep the history.
* **Watch the gaps list** weekly. It is the cheapest quality improvement available.
* **Check what the agent actually retrieved** when an answer looks wrong, using the explainability
panel, before rewriting the entry. Often the entry is fine and a different one is winning.