Skip to main content
The doublePass() strategy performs extraction in two passes: first, it processes chunks in parallel and merges them with an LLM; second, it processes chunks sequentially using the merged result as context to refine the extraction.

​
Usage

import { extract, doublePass } from 'struktur';
import { openai } from '@ai-sdk/openai';

const result = await extract({
  artifacts,
  schema,
  strategy: doublePass({
    model: openai('gpt-4o'),
    mergeModel: openai('gpt-4o'),
    chunkSize: 100000,
  }),
});

​
Configuration

​
model
LanguageModel
required
The AI SDK language model to use for extraction in both passes.
​
mergeModel
LanguageModel
required
The AI SDK language model to use for merging pass 1 results. Typically a more capable model.
​
chunkSize
number
required
Maximum tokens per chunk. Documents are split into batches that fit within this limit.
​
concurrency
number
Maximum number of concurrent extraction tasks in pass 1. Defaults to processing all chunks in parallel.
​
maxImages
number
Maximum number of images per chunk. Useful for controlling vision API costs.
​
outputInstructions
string
Additional instructions to guide the model’s output format or behavior.
​
execute
function
Custom retry executor function. Defaults to runWithRetries.
​
strict
boolean
Enable strict mode for structured output validation. Defaults to false.

​
When to use

  • You need the highest extraction quality
  • You have complex documents requiring multiple perspectives
  • You’re willing to pay for double processing
  • Initial parallel extraction needs refinement from sequential context

​
How it works

  1. Pass 1 - Parallel: Extracts from all chunks concurrently
  2. Pass 1 - Merge: Uses an LLM to merge parallel results
  3. Pass 2 - Sequential: Re-processes each chunk sequentially, using pass 1 results as context to refine and improve extraction

​
Trade-offs

Advantages:
  • Highest extraction quality
  • Combines benefits of both parallel and sequential strategies
  • Second pass can correct mistakes from first pass
  • Fast initial extraction, then refinement
Limitations:
  • Highest token usage (2x extraction + merge)
  • Slowest overall processing time
  • Most expensive strategy
  • May be overkill for simple documents

​
Performance characteristics

The strategy estimates batches.length * 2 + 3 steps:
  1. Prepare
  2. Pass 1: Extract from batch 1 through N (parallel)
  3. Pass 1: Merge
  4. Pass 2: Extract from batch 1 through N (sequential)
  5. Complete

​
Example with concurrency control

import { extract, doublePass } from 'struktur';
import { openai } from '@ai-sdk/openai';

const result = await extract({
  artifacts: complexContractArtifacts,
  schema: contractSchema,
  strategy: doublePass({
    model: openai('gpt-4o'),
    mergeModel: openai('gpt-4o'),
    chunkSize: 100000,
    concurrency: 5,
    maxImages: 10,
    outputInstructions: 'Extract all clauses, dates, and parties with high precision',
  }),
  events: {
    onStep: ({ step, total, label }) => {
      console.log(`[${step}/${total}] ${label}`);
    },
  },
});

console.log('High-quality extraction complete');
console.log(`Total tokens used: ${result.usage.totalTokens}`);

Build docs developers (and LLMs) love

Get started for freeTalk to us