Overview

The Knowledge Bank is the foundation for grounded, accurate agent responses. Instead of relying solely on general model knowledge, an agent connected to a Knowledge Bank draws on the specific business, product, and policy content you provide, so its answers reflect your business rather than a generic model output.

What the Knowledge Bank does

When a knowledge source is attached to an agent, the agent references its content at conversation time to answer user questions accurately and consistently. This allows an agent to:

• Answer questions specific to your products, policies, or services
• Stay consistent with official documentation rather than improvising
• Reduce incorrect or fabricated responses by grounding answers in real content

How processing works

When a document is added to the Knowledge Bank, it does not simply get scanned for text. Each document goes through a structured processing pipeline that reads and understands the document the way a person would.

Processing time varies from a few seconds to a few minutes, depending on the size and complexity of the document.

note

What sets Flam's Knowledge Bank apart Most knowledge base implementations extract raw text and discard everything else. Flam's processing pipeline is layout-aware: it reads the complete structure of a document, not just isolated blocks of text.

During processing, the system:

• Extracts text in its original reading order

• Parses tables, preserving row and column relationships so structured data remains meaningful

• Reads images, charts, and graphs, and generates descriptions of their content

• Understands the overall layout of the document, so information is interpreted in context rather than as disconnected fragments

This means a product brochure with a pricing table, a policy document with an embedded process diagram, or a report with charts can all be understood as a whole, not just as plain text. The agent can then reference facts that live inside a table, or reason about what an image or chart communicates, when forming a response.

This is a core differentiator for Flam Visual Agents against knowledge bases that only support plain-text ingestion.

note

Rich Responses The descriptions generated for images, charts, and tables during processing are also used by Rich Responses, which is live in production. This allows the agent to surface the relevant image, chart, or table alongside its text answer, rather than describing it in words alone.

Where the Knowledge Bank fits into agent creation

Sources are added and managed centrally from the Knowledge Bank. Once a source reaches an Available status, it can be attached to any agent from Build > Capabilities > Knowledge Bank. A single source can be reused across multiple agents, so shared knowledge only needs to be maintained in one place.

For step-by-step instructions on adding, processing, and managing sources, see Managing documents.