Knowledge vs live data

Knowledge in raia Command is where you stage, prepare and organize reference material. Command converts uploads to a readable format, adds metadata, scans for sensitive information, and trains approved content into an agent's vector store. The agent then searches that store to answer questions.

A vector store is good at finding relevant passages. It is not a database. It can't reliably count records, total a column or tell you a value that changed an hour ago.

Learn more: Data Room and Training.

Good sources are approved and relatively stable:

• Help center articles and FAQ pages
• Product manuals and documentation
• Internal procedures, for internal agents only
• Policies and approved answers
• Event programs, agendas and venue information

• Very large PDFs may need splitting.
• Scanned documents need text recognition. Check that the text came through.
• Tables, charts and diagrams should be checked after conversion, because they don't always survive intact.
• Label each document with the product, version or audience it covers, so answers don't mix them up.

Source quality sets answer quality. Time spent here is time saved in testing.

Fetch anything that changes, belongs to one customer, or must be exact:

• Customer accounts, plans and balances
• Contacts, opportunities and pipeline
• Tickets and their current status
• Inventory, transactions and sensor readings

Give the agent live access to the system that owns the data, through a connector, and limit it to the tools it needs. See Connectors: two jobs.

Sometimes the best answer combines the two. A support agent can search imported help articles for a high-quality answer, then check the live version of one article when being current really matters. That second lookup adds time and cost, so use it deliberately rather than by default.

When people ask questions like "how many orders shipped late last month?", the agent needs complete, structured data and real arithmetic:

• For a bounded dataset that doesn't need to be live, use Tables in raia Command. Tables keep exact rows and columns and offer read-only access to agents.
• For live business data, connect to the system through an API or MCP server that returns the right records.
• For calculations, switch on the agent's code interpreter or calculate on the server. Don't rely on the model to do arithmetic in its head.

Code is too large, and changes too often, to retrain on every commit. A hybrid works well: import selected repositories and branches periodically so the agent understands the overall structure, and give it live access to the repository for current files and exact locations. Keep branch and version labels on everything, so answers never blend two versions together.

Command scans content for sensitive data before training. The rules are configurable, and they matter in both directions. A rule that masks something an agent genuinely needs, such as network addresses in a networking guide, will weaken answers. Don't assume a particular value is masked or kept; check the rules configured for that knowledge. Never switch off a security rule for convenience without the data owner's approval.

Learn more: Settings and security and the Trust Center.