Anvia

Analyze a PDF ​

Type: Recipe

Outcome ​

Attach a PDF to a prompt and ask a document-capable completion model for a concise analysis. Use direct document input for small, one-off files; prefer text extraction, OCR, and retrieval for repeated or large document workflows.

Prerequisites ​

  • @anvia/core, @anvia/openai, and a server-side OPENAI_API_KEY
  • A model whose capabilities.documentInput is true
  • A provider-reachable, authorized PDF URL

Implementation ​

ts
import { Agent } from '@anvia/core/agent'
import type { UserMessage } from '@anvia/core/completion'
import { OpenAIClient } from '@anvia/openai'

const model = new OpenAIClient({ apiKey: process.env.OPENAI_API_KEY! })
    .completionModel({
    modelId: 'gpt-5.6-sol',
    api: "responses"
})
if (!model.capabilities.documentInput) throw new Error('Selected model has no document input.')

const analyst = new Agent({
  id: 'document-analyst',
  model: model,
  instructions: 'Use only the attached document. Say when evidence is missing.',
})

const prompt: UserMessage = {
  role: 'user',
  content: [
    { type: 'text', text: 'Summarize the document and list its stated action items.' },
    {
      type: 'file',
      data: { type: 'url', url: 'https://example.com/private-report.pdf' },
      mediaType: 'application/pdf',
      filename: 'report.pdf',
    },
  ],
}

const response = await analyst.generate({ messages: [prompt] })

if (response.type === 'response') {
  console.log(response.output)
}

Run and expected behavior ​

Replace the URL, run pnpm tsx pdf-analysis.ts, and expect a model-generated summary grounded in the PDF. A URL the provider cannot fetch, a file over provider limits, or unsupported document input causes failure.

Boundaries ​

Validate scheme, hostname, MIME type, size, page count, ownership, and malware status before making a URL or upload available. A model summary can omit or misread content and is not evidence that the document is authentic. Do not expose permanent public links to private files.

In production, use short-lived signed URLs, timeouts, retention controls, and document-specific evaluations. Parse once with an application-selected parser or OCR service and index chunks when users will query the same file repeatedly.

Source and extensions ​

Run the PDF attachment cookbook. Next, add citations from extracted chunks or use OCR for scanned pages.

Built for Anvia.