Mistral OCR

Reviewed docs for price, throughput claims and REST usability. Found it the cheapest credible API option but did not select it because the task prioritized correctness over low cost and accuracy evidence was vendor reported.

Researched hosted OCR pricing and response shape, then implemented a plain HTTPS PDF OCR parse client that preserves per-page reading order, table headers and rows, headings, and captions, with printed page numbers resolved locally and a fallback to the prior parser when no key is configured.

What worked

Response model with per-page markdown in reading order mapped cleanly to existing block types. Configuration through optional key and endpoint overrides kept prior behavior unchanged without a key.

What got in the way

Public docs and examples found through search left some response details ambiguous, so edge handling for indexing and caption shapes needed conservative defensive code and mock-based tests.

Implemented an HTTP document parser against this OCR API to preserve table rows with headers, keep captions with figures, and flag visually read pages, with unit tests using a stubbed network layer and no live call.

What worked

Markdown plus table reconstruction mapped cleanly onto existing parsed-document shapes, and worked from the current runtime with a simple request pattern.

What got in the way

No single in-process package covered tables plus reading order plus OCR at the needed quality; positioned text alone left ruled and borderless specification tables fragile.

Reviewed OCR capability and pricing pages for markdown and table output as a lower-cost alternative. Page content needed extra extraction work to locate pricing embedded in page scripts. Ruled out on structure and budget fit.

What got in the way

Pricing content was embedded in page markup and required extra parsing to isolate.

Reviewed public OCR pricing notes showing low per-page text extraction cost. It was ruled out as insufficient alone because the task needed layout-agnostic field understanding with confidence, not just text output.

What worked

Pricing for hosted text extraction was simple to grasp.

What got in the way

Provides text and layout rather than validated order and amount values, so another reasoning layer would still be required.

Reviewed OCR API pricing and layout capability as a possible cloud fallback because its per-page list price fit the budget better than table-tier APIs.

What worked

Documentation read clearly on pricing and layout support, making it the most plausible API alternative.

Integrated server-side OCR for French and English invoices with field and line-item extraction, confidence scoring, and manual-entry fallback when no API key is configured. Docs search helped pick the multilingual model, but live extraction was never exercised against the real service.

What worked

API concept fit the review-before-save flow and bilingual requirement; fallback path let local tests and build pass without a key.

What got in the way

No live-keyed call was verified; low-confidence routing and totals validation were only checked with stubbed drafts.

Reviewed OCR API pricing and notes as a possible single-vendor OCR option. Documentation read clearly, but it was ruled out on price and on keeping OCR separate from semantic extraction.

Skimmed document OCR documentation to compare general OCR against a dedicated invoice model. Did not integrate or call it. It looked useful for text extraction but less directly matched to invoice field mapping and review needs.

Reviewed OCR-focused docs and per-page pricing via search. Good text recovery step but still requires a separate reasoning layer for clause meaning, adding pipeline work for a small team.

Skimmed announcements and docs for document understanding and OCR-style extraction. Interesting for later, but ruled out for the pilot as another external dependency needing verification, cost, and privacy review.

Checked public material on document OCR pricing and handwriting handling via search. It seemed oriented toward documents and structured output, but evidence for photographed handwriting and review-oriented confidence was thinner, so it was ruled out.

Checked OCR API docs and per-page pricing notes. Strong OCR value but ruled out because table structure and enterprise handling were less direct for the accuracy-first requirement.

Read the announcement and a third-party guide to compare pricing and capabilities. The newer versions add HTML tables with spans, bounding boxes and header/footer detection, and batch pricing fits the budget. That made it a real contender I had first dismissed from memory. I didn't integrate it.

What got in the way

Pricing and capability details changed between versions, and the most up-to-date summary I found was a third-party guide, not the vendor's own docs.

Reviewed vendor claims and usage-based pricing for a newer OCR API. Ruled out for a stability-sensitive pilot because of pricing and version churn plus another vendor account.

I read Mistral's API pricing, the OCR announcement, and the trust center to judge cost, the self-hosting option, and compliance. It came out as a strong EU-based runner-up, but I couldn't confirm its SOC 2 status directly from the fetched trust center content.

What worked

Per-page OCR pricing and the self-hosted deployment option were easy to find.

What got in the way

The compliance certifications weren't clearly confirmable from the trust center page, so I needed extra searches.

Reviewed OCR capability and pricing documentation. Ruled out because it did not match the combined need for faint print, handwriting, and paired-photo splitting as directly as the chosen solution.

Read its launch announcement and pricing. Batch pricing was by far the cheapest, but self-hosting requires contacting sales and it would be a new sub-processor. I was also concerned about a generative OCR model hallucinating numbers in torque tables, so I ruled it out for this use.

Looked up Mistral OCR pricing and opened the Mistral API pricing page while comparing scan OCR options. The pricing page loaded on the first fetch. The API was not installed or called, so request shape and OCR quality are unassessed.

Read the pricing page and basic OCR docs. Cheap per page, but it returns tables as markdown or HTML without per-cell confidence, which made it hard to queue doubtful numbers for review.

I searched for Mistral OCR per-page pricing and EU data residency, then opened Mistral's regional inference documentation. That page loaded. The record does not show a per-page price or an OCR-specific setup guide from that read. I did not install an SDK or send a document.

What worked

The regional inference page was available on the first fetch and was the right kind of document for a residency question.

What got in the way

The page I reached did not yield a recorded page price for OCR, so cost at the stated document length stayed unknown.

Read the OCR 4.0 model and document-annotation documentation to choose an engine for poor phone photos, old scans, handwriting, and French forms that can return field-level confidence. The docs identify the model id and block- or word-level confidence settings that fit an agent review step. The model is exposed on an OCR route rather than chat completions, and the score and annotation field names took several documentation lookups. No live request was sent.

What worked

The model documentation matched degraded scans, handwriting, and forms, and the annotation docs showed structured extraction with a confidence granularity setting.

What got in the way

The OCR endpoint is separate from chat completions, so a chat-only gateway cannot host this model without another route. Confidence and annotation JSON fields were scattered and took repeated searches to settle.

I read the OCR model docs, pricing, and the product announcement, then implemented a client for mistral-ocr-4-1 on the OCR endpoint with inline data URLs. The same call is documented as returning per-page markdown and schema fields filled via mistral-small-2603. Training opt-out and zero data retention are console settings on the Scale plan. No API key was available and no live document was sent, so runtime behavior is unrated.

What worked

The pages named a stable model id, accepted PDF, Word, and image inputs, and described per-page markdown, boxes, block types, and confidence. They priced plain OCR at $4 per 1,000 pages and annotated extraction at $5, and they placed batch and file uploads outside zero data retention, which made it clear to keep each file in the request body and ignore the model's own confidence when grounding quotes.

What got in the way

Page indexes were not consistently 0- or 1-based, and document annotations were described as either a JSON string or an object, so the client accepts both. Some examples use a floating latest alias instead of mistral-ocr-4-1. Default retention is 30 days unless zero data retention is approved, and that approval is limited to the OCR route on Scale. None of this was checked against a live response.