Knowledge base

A knowledge base lets your agent answer from your documents instead of only what the model learned during training. You add documents once; afterwards the agent searches them whenever a question needs them, and quotes what it found.

This is useful when the answer lives in material the model has never seen — internal handbooks, product specs, meeting notes, policies, support history.

Knowledge lives in your agent's workspace, alongside the chat. Open your agent from console.ecohash.com → Agents, then choose Knowledge Base in the sidebar — next to Workflows and MCP Tools.

Corpora

Documents are grouped into a corpus — a named collection you search as a unit. A corpus is scoped to your account, so several agents can share one.

Most people start with a single corpus (for example handbook) and add more only when they want searches kept separate — say product-docs and support-tickets, so a product question never pulls in a support transcript.

Add your first document

1. Open Knowledge Base in your agent's workspace.
2. Give the corpus a name, or pick an existing one.
3. Paste the text, or upload a file, and give it a title.
4. Save.

Adding a document is not instant. The text is split into passages and indexed for meaning rather than exact wording, which takes a few seconds for a short document and longer for a large one. The document appears in the list when it is ready to search.

There is no need to format the text specially — plain prose works well. Very long documents are handled automatically; you do not need to split them yourself.

Ask a question

Once a document is indexed, just ask the agent normally:

What does the handbook say about expenses?

The agent decides a search is needed, retrieves the most relevant passages, and answers from them. Because the answer comes from retrieved text rather than memory, it can point at what it used.

If you want to be certain the knowledge base was consulted, ask explicitly:

search the knowledge base for our refund policy and quote it

Memory vs knowledge base

These are two different things and it is worth keeping them apart:

	Knowledge base	Memory
Holds	Documents you add	Short facts the agent chooses to keep
You control it	Yes — you add and remove documents	The agent writes to it during conversation
Good for	Reference material	Preferences, ongoing context
Example	The employee handbook	"This user prefers metric units"

Memory means the agent can carry a detail from one conversation into the next — you tell it once that your team ships on Thursdays, and it still knows next week. Use the knowledge base for anything you would want it to quote.

Managing documents

From the Knowledge Base panel you can:

• add a document to a corpus
• see what a corpus already contains
• remove a document you no longer want searched
• create a second corpus and keep topics apart

Removing a document takes it out of future searches. Answers already given in past conversations are not changed.

Walkthrough: your first knowledge base

Copy these messages into your agent's chat exactly as written. The whole thing takes about three minutes.

Step 1 — create a corpus and add a document. Send this as one message:

Create a knowledge base corpus called team-handbook and add this document to it, titled "Expenses":

Expense claims must be submitted within 30 days of the purchase date. Claims over £200 need approval from your team lead before submission. Travel booked through the company portal does not need a claim — it is billed directly. Receipts are required for anything over £15. The finance team processes claims on the last Friday of each month.

The agent creates the corpus, adds the document, and confirms. It may take a few seconds to finish indexing.

Step 2 — check it landed:

list what's in the team-handbook corpus

You should see the Expenses document listed.

Step 3 — ask a question that is only answerable from that text. Start a new chat first, so you can be sure the answer is not just left over in the conversation:

What's the deadline for submitting an expense claim?

The agent should answer within 30 days of the purchase date, and say it came from your Expenses document.

Step 4 — ask something more specific, to see it retrieve a detail rather than a summary:

Do I need a receipt for a £12 coffee?

The answer should be no — receipts are required over £15. If the agent gets this right, retrieval is working: that threshold appears nowhere except your document.

Make it yours

Now replace the sample with something real:

1. Swap the document. Paste in a page of your own — a policy, a product spec, an FAQ, meeting notes. Same message shape:

   Add this document to team-handbook, titled "YOUR TITLE": YOUR TEXT

2. Ask the questions you would actually ask. The useful test is a question a colleague would ask you, where the answer is a specific detail rather than a general topic.

3. Split by topic when it gets mixed. If answers start pulling in unrelated material, put the next batch in its own corpus:

   Create a corpus called product-docs and add this: …

   Then point the agent at one when you want to narrow it:

   search product-docs for the rate limit on image generation

If the answer looks wrong

• "I don't know" or a vague answer — the document may still be indexing. Ask list what's in <corpus> and confirm it appears, then try again.
• The agent answers from general knowledge instead of your document — say search the knowledge base for … to make retrieval explicit.
• No tools available at all — the agent needs the EcoLink gateway tools enabled; see MCP tools & skills.

Costs

Adding a document and searching it both use platform models — indexing converts your text for search, and a search compares your question against it. These are small charges billed to your wallet balance like any other model usage, and appear in your usage records.

Related

• Agents — creating and running an agent
• MCP tools & skills — enabling the EcoLink gateway tools
• Workflows — using a knowledge search as a step in an automation