Agent Bricks: Knowledge Assistant vs Information Extraction

1.Knowledge Assistant: cited answers over your documents

Agent Bricks gives you ready-made agents for common problems, so you don't have to hand-assemble a retriever, a prompt and an LLM. Knowledge Assistant covers the most common one: people asking questions about a body of documents. It builds a chatbot that answers with citations. It uses an Instructed Retriever approach, which addresses limitations of traditional RAG. The documentation's own examples are answering from product documentation, answering employee HR-policy questions, and answering customers from a support knowledge base. The result is an agent endpoint. You can chat with it in AI Playground or call it from your own applications. Subject matter experts improve it by adding questions and natural-language guidelines in the build experience, so you don't retrain anything by hand.

You can add up to 10 knowledge sources. Each one is one of three types, and each type comes with its own constraints. For every source you also write a "Describe the content" note. This is not just a label: the agent reads it to decide which source to search for a given question.

Sources1

2.Information Extraction: from documents to schema fields

Some requirements aren't about answering questions. They're about turning a pile of documents into rows. Information Extraction does this. You define a schema, and each document or text value becomes a set of structured fields that you can use for analysis, reporting, or downstream agents. The documented examples are legal parties and terms from contracts, line items and payment terms from invoices, and key details from medical records. You can use it from the Agent Bricks UI, in SQL, and through the REST API. Your input data must be in a Unity Catalog volume or table, with at least one file or row.

There are three ways to define the schema. You can describe what you want in natural language and click Generate Schema, which produces a JSON schema with field names and definitions. You can add fields by hand with a name, type and description. Or you can edit the JSON directly. Three optional settings change the output: Precision mode (better accuracy for nuanced fields and long documents), Citations, and Confidence scores. To refine the results, give natural-language feedback on individual inputs, which auto-tunes the field descriptions, or edit the descriptions yourself. Versions let you compare configurations and restore an earlier one.

When the agent performs well enough, click Use Agent. You can either open a SQL query that applies ai_extract with your schema to the whole volume or table, or create a Lakeflow pipeline that runs on a schedule and writes new extractions to a streaming table. Applications call the REST endpoint instead. Its default limit is 120 requests per minute per workspace.

SELECT  ? (
  'Invoice #12345 from Acme Corp for $1,250.00 dated 2024-01-15',
  '{
    "invoice_id": {"type": "string", "description": "Unique invoice identifier"},
    "vendor_name": {"type": "string", "description": "Legal business name"},
    "total_amount": {"type": "number", "description": "Total invoice amount"},
    "invoice_date": {"type": "string", "description": "Date in YYYY-MM-DD format"}
  }',
  options => map('version', '2.1')
);

Sources2

3.Choosing between the two

Both bricks read unstructured documents, which is why they get confused. The deciding question is who consumes the output. When a person asks a free-form question and wants a grounded answer with a source, use Knowledge Assistant. When a table, report or downstream system needs the same fields from every document, use Information Extraction. One gives a conversation over the whole collection. The other gives a schema applied to each document.

Sources1

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.A Knowledge Assistant chatbot is the right tool for pulling invoice totals and dates into a reporting table.Why is that wrong?

   Knowledge Assistant returns cited conversational answers. Information Extraction applies a defined schema to each document and produces structured fields, which you can run at scale with ai_extract or a scheduled Lakeflow pipeline.

   Covered in Information Extraction: from documents to schema fields

2. 2.Any Unity Catalog table of documents can be attached to Knowledge Assistant as a 'Files in a Table' source.Why is that wrong?

   The table must be a streaming table or have Change data feed enabled, and it needs a content column and a metadata STRUCT column. Tables from sources like Jira or Confluence may need to be transformed first.

   Covered in Knowledge Assistant: cited answers over your documents