What Is a Vector Store?
In the previous modules, you gave your agent its brain by writing Instructions. Now you need to give it a memory. You do this by uploading documents into the agent's Vector Store.
A Vector Store is not like a traditional file folder where you search for exact keyword matches. It converts your files into mathematical chunks (embeddings) that the AI can search by meaning and context.
How It Works
- You upload a document (e.g., a product manual PDF)
- Highlight 1: Destination folder
- Highlight 2: Convert with AI
- raia Command converts it to Markdown format
- You send the converted document to the intended agent for training, which starts the processing below
- Highlight 1: Existing fictional Northstar files — not newly processed.
- Highlight 2: Convert: Ready
- Highlight 3: Train: completed
For the full, illustrated sequence — upload, train and check an answer — see the knowledge worked example.
- The Markdown is split into smaller sections (chunks)
- Each chunk is converted into a mathematical vector (embedding)
- Once the file list shows the document as Trained, ask a test question; the agent searches the Vector Store for the most semantically similar chunks
- The retrieved chunks are injected into the context window alongside the system prompt
- The LLM uses both the instructions AND the retrieved knowledge to formulate its answer
This entire process is called Retrieval-Augmented Generation (RAG).
Why Semantic Search Matters
Traditional keyword search fails when users phrase questions differently than the document is written. For example:
| User Question | Document Contains | Keyword Match? | Semantic Match? |
|---|---|---|---|
| "Can I get my money back?" | "Our refund policy allows returns within 30 days..." | No ("money back" ≠ "refund") | Yes (same meaning) |
| "How do I cancel?" | "To terminate your subscription, navigate to..." | No ("cancel" ≠ "terminate") | Yes (same intent) |
The Vector Store searches by meaning, not just words, so it can match questions that use different wording from your documents. Test with the phrasings your customers actually use.
Key Facts About the Vector Store
- Uploading is not the same as training. After you send a document for training, wait until it shows as trained, then ask a test question to confirm the agent uses it
- You choose which agent receives a document when you send it for training, so check the selected agent before you confirm
- For security and data-handling information, see the Trust Center
- If you upload a file with the same name as one already there, raia Command asks what to do: skip it, or replace (overwrite) the existing file. Only a file you choose to replace is updated. Check the converted content, send the updated file to the intended agent for training, and wait for Trained. Ask about a changed fact and compare the reply and Sources with the updated file; conversion alone does not train the agent.
Important: You do not need to manage the vector database infrastructure. raia Command handles conversion, chunking, embedding, and storage for you. You still choose when to send documents for training, and you check the result with a test question.