MIDAAS

MIDAAS · Multimodal Intelligent Document Agentic AI System

Ask your documents. Get the answer and the page it came from.

MIDAAS reads every document your team depends on, including the tables, the wiring diagrams and the fine print, and answers questions in plain language. Every fact is cited to its page. When the question could mean two things, it asks instead of guessing.

Hosted for you at your own address, or run it in your own AWS, Google Cloud, Azure or data center. Pay for what you use, not per seat.

Service library · Ask

Which sensors plug into connector CN44 on the indoor board, and what's the cooling operating range?

Searched Indoor units Service manuals 2024

CN44 takes two thermistors: TH2 (liquid pipe) and TH5 (two-phase pipe)p. 523. The label only appears on the wiring diagram, not in the text.

In cooling, the unit is rated for 67 °F to 115 °F outdoor dry bulbp. 41.

Wiring diagram, p. 523Image sent to the model with the question
Operating range table, p. 41Read from the table row, not a paragraph
Both answers checked against their sources
  1. ConnectS3, Drive, SharePoint, 3,000+ apps
  2. ReadText, tables, diagrams, scans
  3. OrganizeInto collections, automatically
  4. AskIn plain language, in the app or API
  5. RouteTo the right collections, or ask you
  6. AnswerCited to the page, image included

What the name stands for

Six letters, six things it does.

Multimodal Intelligent Document Agentic AI System

Each word is a promise about how your documents are read and how your questions are answered.

M

Multimodal

Reads text, tables, diagrams, photos, scans, audio and video, and shows the model the page image when the answer is in a picture.

I

Intelligent

Understands what's being asked, which product or contract it's about, and which of your collections hold the answer.

D

Document

Built around documents as they really are: chapters, sections, tables and cross-references kept intact, every fact cited to its page.

A

Agentic

A team of agents, each with one job: understand, route, search each collection, check the evidence, and write the answer.

A

AI

Reads every page the way an expert would: describes the diagrams, pulls the facts out of tables, and writes the answer in plain language from the evidence.

S

System

Not a demo. Connectors, sign-in with passkeys, access control, a tamper-evident audit trail, metered billing with spend limits, and monitoring, hosted or in your cloud.

The problem

Search finds documents. Your team needs answers.

A technician on a roof, an engineer checking a spec, an analyst reading a contract: each knows the answer is in there somewhere. Keyword search hands them forty PDFs. A generic chatbot hands them a confident sentence with no page number, sometimes from the wrong model's table.

  • The answer is in a table or a drawing. Most document AI reads the paragraphs and skips the rest. MIDAAS reads tables row by row and looks at diagrams the way a person does.
  • Ten thousand documents blur together. Throw everything into one index and similar products start answering for each other. MIDAAS keeps each product line, client or contract set in its own collection.
  • A wrong answer costs more than no answer. MIDAAS shows its sources, says when it can only answer part of a question, and asks when "Model 123" could mean two things.
  • Your documents live in twenty places. Connect the buckets, drives and apps where they already are. New and changed files are picked up on their own.

How it works

Read once, carefully. Answer many times, with evidence.

Every document goes through the same pipeline once. Each stage is all-or-nothing, so a half-read document is never searchable. Every question then goes through a team of agents that each do one job.

When a document arrives

once per file
  1. Register and de-duplicate

    The same file uploaded twice, or from two places, is stored once. A new version replaces the old one everywhere.

  2. Parse the layout

    Pages, headings, tables, figures and reading order are recovered, from clean PDFs to scanned legacy files.

  3. Sort it into collections

    By a metadata field, a rule you write, or AI classification. One document can belong to several collections without being copied.

  4. Keep the outline

    Every passage knows its chapter, section and subsection, so "the torque spec in Section 4.2" means something.

  5. Look at every image

    Diagrams and photos are described with the whole page in view, and the identifiers printed on them are indexed.

  6. Extract the facts

    Model numbers, table rows, attributes and cross-references go into a structured store, so lookups don't depend on wording.

When someone asks

every question
  1. Understand the question

    Pull out the products, parts and terms. Split a multi-part or comparison question into its pieces.

  2. Pick the collections

    Match the question to the collections that hold its subject. If it fits two equally, ask which one.

  3. Search in parallel

    Meaning-based and keyword search run together in each collection; exact table lookups go to the facts store.

  4. Rank the evidence

    A reranker orders everything found across collections and the best fits are kept within the model's budget.

  5. Answer from the evidence

    The model gets the passages and the actual page images, and answers with a citation on every fact.

  6. Say what's missing

    If part of the question isn't covered by your documents, the answer says so instead of filling the gap.

What it reads, and from where

Any file your team keeps. Wherever it already lives.

No export step and no reformatting. Point MIDAAS at the sources and it keeps up as files are added, changed or deleted.

Documents

PDF, Word, Excel, PowerPoint, HTML, Markdown, LaTeX, CSV, plain text, email, and old .doc, .xls and .ppt files.

Images and scans

PNG, JPEG, TIFF, WebP. Scanned pages, photos of equipment, schematics, exploded views.

Audio and video

Recorded trainings and walkthroughs are transcribed and searchable alongside the manuals.

Structured XML

Patent, scientific-article and financial-filing formats, with their structure kept.

3,000+ apps via connectorsAmazon S3Google DriveSharePointDropboxConfluenceNotionSlackSalesforce FilesYour own MCP serverWebhooksUpload or API

Deploy

Use it hosted, or run it where your data already is.

Both are the same system and the same code. Choose the one that matches how much you want to operate yourself.

Hosted

MIDAAS Platform

We run it. Start a workspace, add your documents, invite your team and start asking.

  • Your own workspace at yourname.midaas.app, with its data kept apart
  • Nothing to install or keep patched
  • Metered billing, itemized per event, with a daily spend limit
  • Settings for people, access, usage and activity

Start a workspace →

Your cloud

MIDAAS Framework

Run it in your own account, on your own terms, next to the data it reads.

  • AWS, Google Cloud, Azure or your own servers
  • Docker, with workers you scale out for bulk loads
  • Postgres-only option (pgvector), no separate vector database
  • You set the rates if you resell it

Built to scale

Add workers for a bulk load of thousands of documents and scale back when it's done. New and changed files are picked up on their own.

Watch it work

Every model call is traced with its prompt, cost and latency. Collection health, ingest jobs, failures and retries are on one screen, with a fix for each.

Security and control

Who sees what is decided per collection.

Built for teams where the contracts folder and the service manuals shouldn't be open to the same people.

  • Roles and permissions. Seven ready-made roles and 22 permissions, with overrides on any single collection.
  • Sign-in without passwords. An emailed link or code, with a passkey or authenticator app required for administrators. Single sign-on (Google, Microsoft Entra ID, Okta) is available on the Enterprise plan, on request. Scoped API keys for services and scripts.
  • A tamper-evident audit trail. Who uploaded, deleted, changed access or signed in, and when, in a chained record you can verify at any time.
  • Your own workspace. Each organization has its own address and its own data. Our support can only come in when you allow it, for as long as you choose, and every visit is recorded and emailed to your administrators.
  • Spend limits. A daily ceiling on AI use per workspace, so a runaway script can't run up a bill.
  • Guardrails on every question. Oversized inputs and known prompt-injection attempts are stopped before they reach a model.
  • No half-finished states. Every ingest stage commits fully or not at all, and every failure has a defined fallback and retry.

How we protect your documents · How MIDAAS uses AI · Privacy

Pricing

Pay for what you read and ask. No seats.

One price per question, whatever it takes to answer it: understanding it, searching every collection, ranking the evidence and writing the cited answer. Every charge is a metered event you can see on your workspace's usage page. Invite your whole team: there are no seat limits. When MIDAAS has to ask which product you meant, that turn is free.

What you useUnitRate
Asking a question, answered with citationsper question, all-in$0.45
Reading text and tablesper passage indexed$0.002
Reading an image or diagram, described by AIper image$0.03
Sorting documents into collections with AIper document$0.002
Storageper 1,000 passages / month$0.25
Image storageper GB / month$0.10

A worked example

A team with 1,000 manuals, each about 150 passages and 12 diagrams, asking 500 questions a working day. Reading the library once costs about $660. After that, about $4,990 a month, almost all of it questions answered.

Larger teams

At higher volumes we agree a flat monthly rate and invoice you. Framework customers running their own copy pay their own cloud and model costs.

Rates shown are the current list rates and may change before general availability.

Developers

Everything in the app is in the API.

A REST API with streaming answers, Python and TypeScript SDKs, webhooks when ingests finish, and a command-line tool that does what the app does.

Run your own test set of questions against your own documents and track answer quality from release to release.

# add a document to a collection
midas-ctl ingest file ./service-manual.pdf --collection indoor-units

# ask, with citations
midas-ctl query run "What plugs into CN44?"

# or over HTTP, streamed as it's written
curl -N https://yourname.midaas.app/api/v1/query \
  -H "X-API-Key: $MIDAAS_KEY" -H "Content-Type: application/json" \
  -d '{"query":"What plugs into CN44?","stream":true}'

# measure answer quality on your own questions
midas-ctl eval run --collection indoor-units

Questions

What people ask before they start.

How is this different from putting our files in a chatbot?

A chatbot reads the text it's given and answers fluently either way. MIDAAS decides which of your collections a question belongs to, reads tables and diagrams as well as paragraphs, sends the model the actual page images, and cites every fact. When it isn't sure which product you mean, it asks.

Does it work on scanned and older documents?

Yes. Scanned pages are read with layout recovery, and legacy .doc, .xls and .ppt files are converted before reading. Very poor scans read less well, as they do for people.

Do our documents train anyone's model?

No. Your documents are used to answer your questions only.

How big a library can it handle?

Collections keep each search space focused, so quality holds as the library grows. Ingest workers scale out for bulk loads, and the same system runs from a few hundred documents to millions of passages.

What happens when a document changes?

Upload the new version from the document's menu. Work already done on unchanged pages is reused. The new version replaces the old one in every collection it belongs to and the old passages stop appearing in answers; earlier versions stay in the document's history. Deleting a document removes it everywhere, including the descriptions and search data made from it.

Are conversations saved?

Yes. Every conversation is saved as you go and can be searched, renamed or pinned. Answers show tables and the cited pages. You can ask by voice, and tell MIDAAS to remember things about you or your project, such as the units you work in.

Can we see how well it's answering?

Yes. The app tracks answer quality against a set of your own questions with known answers, plus cost and speed for every question asked.

See it on your documents

Bring a manual you know well. Ask it the hard questions.

In a 30-minute walkthrough we load a few of your documents and you ask what your team actually asks. You'll see the answers, the citations and the pages they came from.