PDF-to-Markdown API
Brainiall Document Reader engine. $0.001/page. Built for RAG and LLM ingestion.

How we compare

Brainiall Document Reader engine is research SOTA on academic and technical PDFs (~95% structure preservation on multi-column papers). v1.0 calibration vs published competitor pricing and metrics; v1.1 will replace with a 100-document head-to-head benchmark.

vs Mistral OCR 3 (December 2026)

Mistral OCR 3 launched December 2026 at $0.002/page with SOTA quality on tables, figures, and math equations. It is the most disruptive new entrant in the category — and we are honest that on raw OCR quality across complex documents, it likely leads. Here is how S4 (Brainiall Document Reader engine) fits next to it.

• Raw OCR quality: Mistral leads on math/figures/multi-column scientific layouts. S4 (Brainiall Document Reader engine) is solid on technical docs with code, tables, and lists — and ships the layout-aware markdown that downstream RAG pipelines actually want.
• Markdown output: S4 (Brainiall Document Reader engine) returns clean GitHub-flavored markdown preserving headings, tables, lists, math, code blocks. Mistral returns plain text + bounding boxes; you build the markdown converter.
• Self-host option: Production-grade engine (Brainiall Document Reader engine, permissive license). Auditable for regulated industries and airgap-deployable. Mistral is API-only.
• Per-page consistency: Brainiall Document Reader engine handles 1000-page documents without context-window splits. Mistral has practical limits on long inputs.
• Audit trail: Per-call audit DB row with 90-day retention. Mistral is stateless.
• Price: $0.001/page (standard tier) vs $0.002 (Mistral). 2× cheaper.

When to pick Mistral: pure raw OCR with SOTA quality on math/scientific layouts, willing to build your own markdown converter.

When to pick Brainiall: need clean markdown out of the box, self-host capability, audit trail, or 1000-page document handling.

Pricing

Discount derived from quality position vs the closest competitor. 90% off when inferior, 80% off at parity, 50% off when superior.

PAYG: $0.001/page (Brainiall Document Reader engine). HD tier (Brainiall Document Reader engine + Surya OCR refinement for scanned/multilingual/handwritten) is on the v1.1 roadmap — not yet available.

One endpoint, structured output

# Convert PDF to clean Markdown
POST https://api.brainiall.com/v1/document/pdf-to-markdown/base64
 {"pdf": "<base64 pdf>"}

# With page range (skip cover, ToC, etc.)
POST https://api.brainiall.com/v1/document/pdf-to-markdown/base64
 {"pdf": "<base64>", "page_range": "3-50"}

# Markdown-only response (no JSON wrapper)
POST https://api.brainiall.com/v1/document/pdf-to-markdown/base64
 {"pdf": "<base64>", "output_format": "markdown"}

# Response includes structure metadata
# { "markdown": "# Title\n\n...", "metadata": {"pages": 48, "char_count": 12048}, "tier": "fast" }

What Brainiall Document Reader engine does well

• Multi-column layouts: academic papers, magazines, technical reports — preserves reading order.
• Tables: extracted as Markdown tables with header/cell preservation, not flattened text.
• Math + code: equations rendered as LaTeX inline; code blocks preserved with monospace fences.
• Headings + structure: H1/H2/H3 hierarchy detected from font sizes + position cues.
• Footnotes + references: linked at-paragraph, not dropped.
• Multilingual: HD tier adds Surya OCR for non-English PDFs and scanned documents.

Built for RAG pipelines

Most PDF parsers output JSON or HTML — your RAG pipeline then has to re-flatten it back to text-with-structure. Brainiall Document Reader engine outputs Markdown directly, which embedders (OpenAI, Voyage, Cohere) handle natively. Cut the conversion step.

Latency profile — two regimes

PDF conversion runs in one of two modes, picked automatically from the input. Plan your integration around the right one — surprises here are the most common reason people call us.

• Native-text PDFs (the fast path): if the file already carries an embedded text layer (most exports from Word, LaTeX, Google Docs, modern report generators), conversion runs at ~1.3 s/page. A 5-page text PDF lands in roughly 6.3 s end-to-end. Synchronous request/response is fine here.
• Scanned or image-only PDFs (the OCR path): if the file has no extractable text (camera scans, fax output, photos of paper, older archives), the document model has to run OCR page-by-page on CPU. Budget ~20–30 s/page. A 5-page scan takes 1.5–2.5 minutes.
• Recommended pattern for scanned input: treat the request like a job, not a synchronous call. Submit, hold the connection or queue the work server-side, render a progress indicator. The same async-job pattern that dubbing and speech-to-speech use applies cleanly to long scanned-PDF conversions.

Press kit & resources

Everything reviewers, integrators and procurement teams typically ask for.

More specialty APIs

Same single API key, same usage-based pricing, different problem solved.

Background RemovalAzure Image Analysis 4.0 alternative

Audio EnhancementDenoise, isolate, cleanup, master

Speaker DiarizationBrainiall Speaker ID engine, who said what

Agent MemoryMem0 alternative, 1/5th the cost