Document Q&A with Citations

The user asks a question against a document. The pipeline extracts only the relevant passages, then answers using just those passages — keeping the answering block's context small and the citations grounded.

Tree

Flow: "document-qa"
├── Block: "extract-relevant-passages" (llm)
└── Block: "answer-with-citations" (llm)

Blocks

extract-relevant-passages (llm)

Below is a document and a question. Return up to 5 passages from the document that are directly relevant to answering the question. Each passage should be quoted verbatim and include its location (paragraph number).

Question: {{parameters.question}}

Document:
{{message}}

Output schema:

{
  "type": "object",
  "properties": {
    "passages": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "paragraph": { "type": "integer" },
          "text": { "type": "string" }
        }
      }
    }
  }
}

answer-with-citations (llm)

Answer the question using ONLY the passages below. Cite each claim by paragraph number in square brackets, e.g. [3]. If the passages don't contain the answer, say so explicitly.

Question: {{parameters.question}}

Passages:
{{previous_output.passages}}

Multi-document variant

If the question runs against multiple documents, wrap the extraction in a loop:

├── Loop over "parameters.documents":
│     └── Block: "extract-relevant-passages" (llm)
└── Block: "answer-with-citations" (llm)

Why this shape?

• Two blocks, not one, because the answering block's job is grounded synthesis — and synthesis quality drops sharply when the model also has to hunt through irrelevant text. The first block does the hunting; the second only sees the small set of relevant passages.
• The full document never reaches the answering block. This is the entire point of the decomposition — and it's also where the token savings come from. A 50,000-token document becomes a 1,000-token passage list.
• Citations are constructed from the structured passage list, not generated freely. Because each passage carries its paragraph number, the answering block can cite by index instead of inventing locations.