MIDAAS · Early access

Document AI that keeps the images, the hierarchy and the honesty.

Multimodal Intelligent Document Agentic AI System. Ingest any file type, preserve chapters and sections, index images as first-class retrieval objects, and answer through agents that route, ask, and admit partial answers.

Request early access

Why another document system

Most RAG stacks summarize the diagram away, flatten the document, and let one index saturate as the corpus grows. MIDAAS was designed around those three failures.

Images are first-class nodes.

The actual image is delivered to the model at query time, not a caption of it.

The orchestrator asks rather than guesses.

Ambiguous questions get a clarifying question, not a confident wrong answer.

All pipeline stages are transactional.

Partial ingest states cannot exist.

Every failure mode has a defined fallback.

The system degrades gracefully and says so.

Any file type

PDF, Office, HTML, Markdown, LaTeX, CSV, images, audio, video, code, email and legacy formats, through an 11-stage pipeline.

Collections against saturation

Content is organized into named collections so one giant index never dilutes retrieval. One physical copy, many logical collections.

3,000+ connectors

Source documents from SaaS systems via MCP, from S3 and blob stores, or from your own MCP servers.

Three-store architecture

Relational facts and relationships, a vector store, and a graph store when the question needs it. Answers come with an evidence matrix.

Enterprise controls

Seven roles, 22 permissions, SSO for Google, Okta and Azure AD, audit log, webhooks, dead-letter queue, Python and TypeScript SDKs.

Measured, not assumed

An evaluation harness with golden datasets, tracing on every call, and a planned answer-validation stage that refuses on weak grounding.

Offerings: MIDAAS Framework, an open developer library you run. MIDAAS Platform, a fully managed hosted service. Same codebase.
Pricing: Consumption only. Metered per chunk, image, caption, routing turn, rerank and answer token. No seat licenses.
Example: 10,000 documents and 5,000 queries a day: about $1,500 a month ongoing
Models: Claude by default, swappable to GPT, Gemini, Bedrock or local models
Deploy: Self-hosted, AWS, GCP or Azure
Status: Early access. 700+ automated tests. Release hardening in progress.

Validated on HVAC service manuals with wiring diagrams and on medical-equipment electrical drawings.