Mixedbread Parsing

Mixedbread Parsing

Parse documents, extract structured content, and run OCR using the Parsing API. Supports PDFs, Word documents, PowerPoint presentations, and images.

Docs: https://www.mixedbread.com/docs/parsing/overview.md Agent-readable docs: https://www.mixedbread.com/docs/llms.txt Latest docs search: https://www.mixedbread.com/question?q=parsing&section=docs

Setup

pip install mixedbread          # Python
npm install @mixedbread/sdk     # TypeScript

export MXBAI_API_KEY=your_api_key

Quick Start

Python:

from mixedbread import Mixedbread

mxbai = Mixedbread()

# Upload and parse a document (waits for completion)
job = mxbai.parsing.jobs.upload_and_poll(
    file=open("report.pdf", "rb"),
    return_format="markdown",
)

for chunk in job.result.chunks:
    print(chunk.content)

TypeScript:

import Mixedbread from '@mixedbread/sdk';
import fs from 'fs';

const mxbai = new Mixedbread();

const job = await mxbai.parsing.jobs.uploadAndPoll(
    fs.createReadStream('report.pdf'),
    { return_format: 'markdown' },
);

for (const chunk of job.result.chunks) {
    console.log(chunk.content);
}

Decision Tree

• Which convenience method?
  • File on disk → upload_and_poll() (uploads + creates job + polls)
  • File already uploaded via Files API → create_and_poll() (creates job + polls)
  • Need async control → upload() or create() then poll() separately
• Which parsing mode? (default is high_quality)
  • Born-digital PDF (selectable text) → pass mode="fast" explicitly. Fastest, lowest cost. Extracts text, structure, and layout.
  • Scanned document, image, or complex layout → high_quality mode (the default). Uses OCR. Extracts text with confidence scores and per-element bounding boxes, handles rotated/skewed pages, multi-column layouts.
• Which return format? → return_format: markdown (default), html, or plain
• Need specific elements only? → Set element_types to reduce processing time

Supported File Types

PDF (.pdf), Word (.doc, .docx, .dotx, .docm, .dotm, .odt, .rtf), Slides (.ppt, .pptx, .ppsx, .ppam, .pptm, .potm, .ppsm, .odp), Images (.jpeg, .png, .webp, .avif).

Element types: text, title, section-header, header, footer, page-number, list-item, figure, table, form, footnote. (Legacy values picture, caption, formula, page-header, and page-footer are accepted but normalized to figure/text/header/footer.)

Each extracted element carries type, content, page, confidence (0–1), and bbox — the bounding box [x1, y1, x2, y2] in page pixel coordinates. Use bboxes to map OCR output back to its location on the page (e.g. evidence highlighting).

Chunking: chunking_strategy defaults to page (currently the only strategy) — one chunk per page.

Workflows

Extract Tables from Documents

Filter for table elements to pull structured data from reports.

Python:

job = mxbai.parsing.jobs.upload_and_poll(
    file=open("financial-report.pdf", "rb"),
    element_types=["table"],
    return_format="html",
    mode="high_quality",
)
for chunk in job.result.chunks:
    for element in chunk.elements:
        if element.type == "table":
            print(f"Page {element.page}, confidence {element.confidence:.2f}, bbox {element.bbox}")
            print(element.content)

TypeScript:

const job = await mxbai.parsing.jobs.uploadAndPoll(
    fs.createReadStream('financial-report.pdf'),
    { element_types: ['table'], return_format: 'html', mode: 'high_quality' },
);
for (const chunk of job.result.chunks) {
    for (const element of chunk.elements) {
        if (element.type === 'table') {
            console.log(`Page ${element.page}, confidence ${element.confidence.toFixed(2)}, bbox ${element.bbox}`);
            console.log(element.content);
        }
    }
}

Batch Parse Multiple Files

Upload multiple files asynchronously, then poll all jobs:

Python:

import os

jobs = []
for filename in os.listdir("./documents"):
    if filename.endswith(".pdf"):
        job = mxbai.parsing.jobs.upload(
            file=open(f"./documents/{filename}", "rb"),
            return_format="markdown",
        )
        jobs.append(job)

# Poll all jobs
for job in jobs:
    completed = mxbai.parsing.jobs.poll(job_id=job.id)
    print(f"{completed.filename}: {len(completed.result.chunks)} chunks")

TypeScript:

import { readdirSync, createReadStream } from 'fs';
import path from 'path';

const files = readdirSync('./documents').filter(f => f.endsWith('.pdf'));
const jobs = await Promise.all(
    files.map(f => mxbai.parsing.jobs.upload(
        createReadStream(path.join('./documents', f)),
        { return_format: 'markdown' },
    )),
);

// Poll all jobs
for (const job of jobs) {
    const completed = await mxbai.parsing.jobs.poll(job.id);
    console.log(`${completed.filename}: ${completed.result.chunks.length} chunks`);
}

Rules

CRITICAL

• Don't double-parse. Store uploads auto-parse documents. Files uploaded with parsing_strategy: "high_quality" automatically get OCR text (images), summaries (images), and transcriptions (audio & video) extracted. These are available as fields on search result chunks. There is no benefit to also running the Parsing API on the same file. Use the Parsing API only for standalone document extraction outside of stores.
• Use upload_and_poll() / create_and_poll() instead of manual polling loops. These methods handle backoff automatically. Manual while loops with retrieve() are fragile and waste API calls.

HIGH

• Specify element_types when you only need certain elements. Requesting all types increases processing time and response size. If you only need tables, set element_types to table only.
• Use fast mode for born-digital PDFs. The high_quality mode adds OCR overhead that provides no benefit when text is already selectable.
• Check confidence scores on OCR output. Low-confidence elements (< 0.5) may contain errors. Filter or flag them.

MEDIUM

• Check job.error before retrying failed jobs. Common causes: unsupported file type, corrupt file, file too large. Blindly retrying wastes quota.
• Use content_to_embed for embedding pipelines. Each chunk provides both content (full text) and content_to_embed (optimized for embedding). Use the latter when feeding into vector stores outside Mixedbread.
• Verify file format before parsing. Only PDF, Word, PowerPoint, and images are supported. Convert other formats first.

Troubleshooting