Skip to main content
A Knowledge Store is a vector-indexed collection of your documents (PDFs, DOCX, CSV, or scraped websites). Once you attach a store to an agent, every user message triggers a retrieval pass — the most relevant chunks are pulled into the prompt and the agent answers from them, with citations. This guide walks you through creating a store, uploading content, and wiring it to an agent — entirely from the dashboard.

​
Prerequisites

  • A signed-in dashboard account at app.agntix.ai.
  • An existing agent (see Build your first agent). You’ll attach the store to it in Step 4.
  • The files you want to index (PDF, DOCX, or CSV), or a list of public URLs to scrape.

​
Step 1: Open the Knowledge Store page

From the left navigation, click Knowledge Store. You’ll land on a two-column layout — list of stores on the left, details for the selected store on the right.
Empty Knowledge Store page

The Knowledge Store page. Click the + button (top of the left panel) to create your first store.

Click the + button at the top of the left panel.

​
Step 2: Create the store

A modal opens with a single required field — Name. Give the store a descriptive name (8–50 characters). Examples: Help Center, Product Docs FY26, Onboarding playbook.
Create store modal

Create Knowledge Store modal — name only. Embeddings, chunking, and the vector index are configured automatically.

Keep stores narrow. Retrieval quality is much higher with one store per topic than one giant store containing everything your company has ever written. You can attach different stores to different agents.
Click Create. The new store appears in the left panel; select it to open the detail view.

​
Step 3: Upload documents

In the store detail view, click the Add dropdown in the top right. Two sources are supported:
  • Upload file — PDF, DOCX, CSV (multiple files allowed in one upload).
  • Add website — scrape a single URL or a list of URLs.
Add dropdown

Add dropdown — pick file upload or website scrape.

Select Upload file and pick one or more documents. Each file is queued and processed asynchronously:
  1. Uploading — bytes transfer to the gateway.
  2. Processing — Agntix chunks the document (~1k tokens per chunk), embeds it (Cohere by default), and writes vectors to Qdrant.
  3. Ready — the document is searchable.
Upload in progress

Upload progress. Status badges flip from Uploading → Processing → Ready.

Wait for every document to reach Ready before testing. Processing time scales with document size (typical: 1–2 seconds per page).
Documents ready

Documents ready for retrieval.

​
Step 4: Attach the store to an agent

Attaching a knowledge store from the dashboard is not available in the current release — the agent’s in-dashboard Knowledge tab is temporarily disabled while it’s being finalized. For now, attach the store over the API as shown below; the dashboard flow will return in a future release.
Copy the store’s ID from the Knowledge Store detail view (or from the create-store API response), then PATCH your agent to enable RAG and point it at the store:
Retrieval behaviour — search type, Top K / Top N, Agentic RAG, and Rerank — defaults to sensible values. To tune it, set the matching config fields; see the Agents section of the API Reference for the exact field names and ranges.

​
Step 5: Test with a RAG-grounded question

Click Test Agent in the page header. Ask a question whose answer is in your uploaded documents — something specific enough that a generic LLM wouldn’t know it.
Test RAG response

The agent answers from the knowledge store. Source snippets are shown inline (or in the response payload for API users).

Compare against the same question with the store detached — the difference should be obvious. If the agent ignores the store, see Troubleshooting below.

​
Verify it works

  • Every uploaded document shows Ready in the store detail view.
  • The agent’s config shows ragEnabled: true with your knowledgeStoreId (check via GET /v1/chat/agents/{id}).
  • Test Agent answers contain facts that are only in your documents.
  • The session log (Analytics tab or /logs-history) shows the retrieved chunks per message.

​
Best practices

One topic per store

Narrow stores out-perform giant stores. Split by product, region, or audience.

Refresh on a schedule

Re-upload changed documents. Old chunks are replaced atomically, no downtime.

Always rerank

Cohere reranking adds ~150 ms but cuts hallucinations dramatically.

Cite sources in the prompt

Tell the agent to cite chunk titles. Builds user trust and makes wrong answers easy to debug.

​
Troubleshooting

​
API alternative

Same flow over HTTP — useful for CI / bulk ingestion / nightly refreshes:
For the full upload, document-listing, and chunk-level endpoints, see the auto-generated Knowledge Store section under API Reference.