space ocr

Product Introduction

1. Definition: Space OCR is a self-verifying Optical Character Recognition (OCR) API and application platform. It is a technical infrastructure service that converts unstructured document images (like photos of receipts, invoices, and forms) into structured, queryable, and verifiable data.
2. Core Value Proposition: It exists to solve the critical problem of trust in automated document processing. Unlike standard OCR APIs, Space OCR provides verifiable data extraction by anchoring every extracted data point to its exact source coordinates on the original image, flagging discrepancies, and eliminating hallucinations. Its core value is delivering a searchable, verifiable database from any document without requiring users to manage backend infrastructure.

Main Features

1. Self-Verifying OCR Pipeline: This is the core technical engine. It performs a multi-stage verification check on every extracted value. First, standard Vision OCR generates character-level coordinates. Then, a multimodal AI model interprets the page content. Each value the model returns is "anchored" to the OCR text at specific vertices. The system cross-checks the model's output against the raw OCR in those anchored regions. Discrepancies trigger a re-read of a cropped image section. The final output includes verification flags (text_verified, needs_review) and the precise vertices for every data point, providing built-in data quality assurance.
2. Multi-Format Output Endpoints: The API offers structured output tailored for different downstream uses. The /ocr/fields endpoint allows users to define specific named fields (e.g., total_amount, date) and returns typed, structured JSON with line items unfolded into rows. The /ocr/markdown endpoint preserves document semantics (headings, lists, tables) in Markdown format for wikis or LLM contexts. The /ocr/text endpoint provides clean, paragraph-structured plain text for indexing and search.
3. Persistent, Queryable Data Space: Beyond one-off API calls, Space OCR provides a built-in data layer. Users can POST /create named folders and sheets. Uploaded documents become persistent rows that stack within these sheets. Users can then GET /view to query this accumulated data with server-side filters (where, sort, select), and GET /space to perform full-text search across folder names, file names, memo bodies, and extracted cell values, creating a ready-to-use document database.

Problems Solved

1. Pain Point: The "black box" problem of AI document processing, where users cannot verify if an extracted value (like an invoice total or a part number) is correct or a hallucination. Traditional OCR and document AI services lack transparent, pixel-level evidence for their outputs, creating risk and manual review overhead.
2. Target Audience: Developers and Engineers building document-heavy applications (fintech, logistics, legal tech); Business Operations Teams (AP/AR, compliance, administration) dealing with receipts and forms; Researchers and Analysts processing surveys or archival documents; AI Agent Developers using MCP (Model Context Protocol) to give agents document-handling capabilities.
3. Use Cases: Automating accounts payable invoice data entry with auditable proof; digitizing and querying a backlog of paper forms or receipts for expense reporting; extracting structured data from product spec sheets for a catalog; enabling AI agents in Claude or ChatGPT to reliably file documents and retrieve specific information later.

Unique Advantages

1. Differentiation: Versus generic OCR APIs (Google Vision, AWS Textract), Space OCR adds the verification layer and persistent queryable storage. Versus other document AI platforms, its focus on coordinate-level verifiability and the built-in no-database data workspace is unique. It provides evidence, not just extraction.
2. Key Innovation: The "anchor and cross-check" verification pipeline is the key technological innovation. By forcing the AI model's output to be grounded in pre-existing OCR coordinates and then running a discrepancy check, it directly tackles hallucination. The integration of this verified output into a native queryable data space (via folders and sheets) means the product is both an extraction engine and a simple application database, reducing system complexity for users.

Frequently Asked Questions (FAQ)

1. How accurate is Space OCR compared to other OCR services? Space OCR measures accuracy not just on text recognition but on value verification. Internal metrics show that cells flagged with text_verified: false are 6.4x more likely to contain a real error, and 92.1% of returned coordinate boxes contain the exact value spelled in the original page's OCR. This transparent, evidence-based scoring is fundamentally different from opaque accuracy claims.
2. Can I use Space OCR to process documents in batch automatically? Yes. The API supports asynchronous processing via a /jobs endpoint and can deliver results via HMAC-signed webhooks, making it suitable for high-volume, automated batch processing pipelines for invoices, receipts, or forms without manual intervention.
3. What is the "no database" feature, and how does it work? The "no database to stand up" advantage means Space OCR includes a built-in data persistence and query layer. Instead of extracting data and then managing your own database (like PostgreSQL), you create a folder and sheet within Space OCR. Each uploaded document appends a row to that sheet. You then use the native GET /view endpoint with filter parameters to query your historical data, eliminating the need for separate database infrastructure for simple applications.
4. How does the MCP (Model Context Protocol) server integration work? The Space OCR MCP server allows AI agents (in Claude, ChatGPT, Cursor, etc.) to interact with the platform as a tool. An agent can programmatically space_create folders, space_upload photos to become rows, and space_view to query stored data. This turns Space OCR into the permanent memory and filing system for AI agents handling document-based tasks.
5. What happens when the verification pipeline flags a value for 'needs_review'? The value is still returned in the JSON output but with the needs_review flag set to true. The vertices for that value are also provided. This allows developers to build workflows that route low-confidence extractions for human review in a UI (like Space OCR's own console, where you can hover over the value to see the source image region) or to apply custom business logic, ensuring data quality without rejecting the entire document.