# Give your AI agents a knowledge base your team controls

What to put in it, which files agents can read, and how access tags decide which documents each person's agents can search

Oct 23, 2026 · Guides · 9 min read · https://arclyx.ai/blog/ai-agent-knowledge-base

## Summary

An AI agent knowledge base is the set of documents your agents search before they answer or act: your bid profile, content brief, pricing policy and handbook. In Arclyx you upload them once and tag each file with who may read it. Every agent then searches only what the person it runs for can read, and cites the file it used. This guide covers what to upload, which file types work, how the access rules behave, and what changes when you share an agent with the whole team.

## Key Takeaways

- **Upload what a job depends on.** The facts an agent would otherwise guess: your bid profile, content brief, competitor notes or pricing policy.
- **PDF, Word, Excel, CSV and text.** Agents search documents by meaning and cite the file; spreadsheets can also be read as tables.
- **Tags decide who reads what.** Members read files that carry a tag they hold. Owners and admins read every shared file.
- **An agent reads as its person.** It never searches a document the person it runs for can't read, and in a Slack channel private files stay out.
- **No tag means private.** A file uploaded without a tag stays private to the person who uploaded it, and can't be tagged later.
- **Sharing an agent doesn't share documents.** Each person runs a shared agent with their own access, and their runs stay private to them.

## What is an AI agent knowledge base?

An AI agent knowledge base is the collection of your own documents that an agent searches for facts it doesn't have: policies, profiles, briefs, price lists and handbooks. The agent reads them at run time, so replacing a file changes the next answer. For a team, the hard part isn't search. It's deciding who can read what.

Retrieval is one of the basic building blocks of an agent. Anthropic describes the starting point of any agentic system as a model "enhanced with augmentations such as retrieval, tools, and memory" [Source 2]. A knowledge base is the retrieval part, filled with your company's facts instead of the open web.

That's also where the risk sits. Once several people share one store of documents, an agent that searches everything can hand a pricing sheet to someone who was never meant to see it. OWASP's list of risks for LLM applications names this directly and asks for "permission-aware vector and embedding stores" [Source 1]. The rest of this guide is about doing that without making it a project.

## What should go in it?

Put in the documents a recurring job depends on, written down once so the agent stops guessing. Good candidates are short, stable and specific to your business: a profile of what you bid on, a content brief, notes on each competitor, a pricing or discount policy. Leave out anything that changes daily or lives better in another tool.

Several of the agents on our [AI agents for business](https://arclyx.ai/agents) page start by reading one document:

| Agent | The document it reads first | What it uses it for |
|---|---|---|
| RFP finder | Your RFP profile | Scope, value range, closing-date rules and exclusions |
| Blog drafter | Your content brief | Topic, audience, angle and the points it must cover |
| Competitor watch | Your competitor notes | What you already know, so it keeps only what's new |
| Handbook answerer | Your employee handbook | HR, IT and policy answers that quote the section used |

What to keep out:

- **Things that change every day.** A live pipeline or a shared drive of working files is better reached through a connector. Agents can search and read Google Drive files and Notion pages with your own sign-in; the [AI agent integrations page](https://arclyx.ai/product/connectors) lists every one.
- **Secrets.** Passwords, API keys and bank details don't belong in a document an agent reads, whoever can see it.
- **Anything you wouldn't stand behind.** OWASP's advice is to accept data "only from trusted and verified sources" [Source 1]. An agent treats an outdated price list as the truth.

## Which file types can agents read?

Arclyx reads PDF, Word, Excel, CSV and plain text files. Documents are searched by meaning, so an agent finds the clause about discounts even if the page never uses the word, and its answer cites the file it came from. Spreadsheets and CSV files can also be read as tables, so an agent can check the columns and query the rows.

| File | Extensions | How agents use it |
|---|---|---|
| PDF | .pdf | Searched by meaning, cited by file name |
| Word | .docx, .doc | Searched by meaning, cited by file name |
| Text | .txt | Searched by meaning, cited by file name |
| Excel | .xlsx, .xls | Searched, or read as a table: columns, rows, queries |
| CSV | .csv | Searched, or read as a table: columns, rows, queries |

Tables are how Finance desk, the month-end agent, checks its numbers: it works from CSV exports uploaded as tables, with no accounting integration to set up first. How many documents you can keep depends on your plan.

## How do you decide who can read which document?

With access tags. Admins make the tags, such as sales, finance or everyone, give each member the tags they need, and tag files when they upload them. A member reads every file that carries at least one of their tags. Owners and admins read every shared file. A member with no tags reads only their own private files.

![The Upload files screen in Arclyx with example data: two files, Pricing policy 2026.pdf and Discount approvals.docx, and the sales and finance tags picked, with the note "People with sales or finance can read them."](https://arclyx.ai/blog/ai-agent-knowledge-base/img/upload-with-tags.6a0bc8c6-1600.webp)

When an admin uploads files, the tags they pick are the files' readers. Pick none and the files stay private to the admin. Members can upload too, but their files are always private to them: only an admin can share a file with a tag.

![Settings, People in Arclyx with example data: the owner and an admin on All tags, a member with sales and support, another with no tags yet, and the access editor open on the member's tags.](https://arclyx.ai/blog/ai-agent-knowledge-base/img/member-access-tags.2a857e46-1600.webp)

Who reads what, by role:

| Who | Reads | Uploads |
|---|---|---|
| Owner or admin | Every shared file in the organization, plus their own private files | Tagged files, or private ones |
| Member | Files with a tag they hold, plus their own private files | Private files only |
| Member with no tags | Their own private files only | Private files only |

Private really means private. A file uploaded with no tag stays out of everyone else's search, admins included, and it can't be tagged later. To share it, delete it and upload it again with a tag. Deleting a file stops Arclyx reading it for everyone who could read it, and it can't be undone.

![The Documents page in Arclyx with example data: five files with the tags that decide who can read each one, a filter by who can read it, and one text file marked Private.](https://arclyx.ai/blog/ai-agent-knowledge-base/img/documents-by-tag.46ba9348-1600.webp)

## Do agents see documents their owner can't?

No. An agent searches documents as the person it runs for: their role and their tags, never more. A run that a schedule or webhook starts reads with its owner's access at most. Two people running the same agent can get different answers, because each search only reaches the files that person can read.

One case is narrower on purpose. When someone asks Arclyx in a Slack channel, everyone in the channel can read the reply, so private files stay out of that answer, the asker's own included. Shared files still count, by the asker's own access. That's why [Handbook answerer](https://arclyx.ai/agents/hr-assistant), which answers policy questions when someone @mentions it in a channel, needs its handbook shared with everyone by an admin. The [Slack AI agent guide](https://arclyx.ai/blog/slack-ai-agent) covers channels and threads in more detail.

Access tags sit with the other limits Arclyx enforces outside the model. Text in a document can't widen what an agent may read, any more than it can pick a different Slack channel. The [AI agent guardrails page](https://arclyx.ai/product/guardrails) lists them all, and the [AI agent security page](https://arclyx.ai/security) sets out who can see which documents and runs.

## How do you share an agent with the whole team?

An admin opens one of their own agents, goes to its Config tab and turns on Share with everyone. A confirm lists what changes. Everyone in the organization can then run it and chat with it, each with their own tools and sign-ins, so each person's run reads only the documents that person may read.

![Sharing an agent in Arclyx with example data: the confirm for sharing Competitor watch with everyone in Acme, listing that everyone can run it, each person runs it with their own tools and sign-ins, runs stay private, and it can't be made personal again.](https://arclyx.ai/blog/ai-agent-knowledge-base/img/share-agent-with-everyone.d8fb7325-1600.webp)

What the confirm tells you, in the app's words:

- **Everyone can run it and chat with it.** Admins can read its instructions; members get the agent without them.
- **Each person runs it with their own tools and sign-ins.** Sharing the agent shares the job, not your Gmail, your HubSpot or your documents.
- **Runs stay private to whoever started them.** An admin manages the agent but doesn't receive copies of what other people put through it.
- **Your schedules and webhooks on it stay yours.**
- **It's one way.** A shared agent can't be made personal again, only deleted, so nobody loses an agent they rely on.

## Knowledge base best practices

Keep each document short, current and about one thing, tag it by team rather than by person, and check that agents find it before they run on their own. These habits come from how Arclyx enforces access and from OWASP's guidance on retrieval: least access, trusted sources and a record of what was read [Source 1].

1. **One document per job.** An RFP profile titled "RFP profile" is easier to find, and to check, than page 14 of a sales deck.
2. **Tag by team, not by person.** People change roles; a finance tag outlives them. Give each member the tags their work needs and no more.
3. **Replace, don't pile up.** When a policy changes, delete the old file and upload the new one, so no agent cites last year's prices.
4. **Use everyone sparingly.** A tag everyone holds is a shared file in all but name. Keep it for handbooks and policies meant for the whole company.
5. **Test the search before it runs alone.** Try it on a draft is a dry run with your real tools: each round shows the tool and what it asked for, so you can see which document it searched.
6. **Read the record.** Every run keeps the documents and pages it searched or opened as tool calls you can open, which is the retrieval log OWASP asks for [Source 1].

Text inside a document can also try to give an agent orders, which is one more reason to upload only files from sources you trust. To write the agent that reads these files, start with [how to write AI agent instructions](https://arclyx.ai/blog/ai-agent-instructions).

## Sources

1. OWASP Gen AI Security Project — *LLM08:2025 Vector and Embedding Weaknesses*, "Prevention and Mitigation Strategies" (checked 23 October 2026). [https://genai.owasp.org/llmrisk/llm082025-vector-and-embedding-weaknesses/](https://genai.owasp.org/llmrisk/llm082025-vector-and-embedding-weaknesses/)
2. Anthropic — *Building effective agents* (19 December 2024), "Building block: The augmented LLM". [https://www.anthropic.com/research/building-effective-agents](https://www.anthropic.com/research/building-effective-agents)