The multimodal warehouse · Documents
Get the answer out of every PDF and scan, with the page it came from.
Contracts, filings, manuals and scans are split into the paragraphs, tables, forms and figures they are made of, each graded for how well it was read. They live in the same warehouse as your video, images and audio, so a manual section can be checked against the training video that teaches it.
From $25/mo. Index once. Query every way.
| paragraph | “Either party may terminate on 60 days notice” | page 7 |
| table | Fee schedule, kept as rows and columns | page 12 |
| form | Signature block with names and dates | page 31 |
| handwriting | A margin note, re-read by a vision model | page 9 |
| page image | An image embedding of the whole page | page 12 |
One warehouse
Your documents sit beside every other file you own.
Each file is broken into rows that point back to the file they came from. Video scenes, images, audio segments and document blocks live in one namespace, so a question can start in one kind of file and finish in another.
Videoscenetranscripton-screen textfacekeyframe Imagesembeddingfacelabelmetadata Audiosegmenttranscriptsoundmetadata
Documentsparagraphtableformhandwritingpage image
Joins your documents make
Which training videos still show the screens this manual changed?
manual sectionsvideo scenes
Runs on the Cross compare stageAttach the vendor record from our database to every contract.
contractsyour SQL table
Runs on the SQL lookup stageLabel every filing by type and sensitivity from our own list.
filingsyour taxonomy
Runs on the Taxonomy enrich stageWhat you can ask
Ask in plain words, or with a file.
› termination for convenience clauses
The paragraphs, with their page
› fee tables over $50,000
Tables matched and filtered
› this page from another scan
Pages that look like it
› unsigned agreements from 2024
Semantic matches filtered by your fields
Start in one click
Point it at your documents
An empty namespace
Connect the bucket your documents already live in and pick the extractors. Studio indexes it and your first query runs against your own files.
Create a namespace →In code
One call to search what was indexed.
The extractors that turn documents into rows:
- Document graph →Paragraphs, tables, forms and figures with a confidence grade
- Image →Page embeddings for layouts text misses
- Text →Multilingual embeddings for every block
from mixpeek import Mixpeek
client = Mixpeek(api_key="YOUR_API_KEY")
# Ask the documents; get back the blocks that answer, with their page
results = client.retrievers.execute(
"ret_YOUR_RETRIEVER_ID",
inputs={"query": "termination for convenience"},
filters={"AND": [
{"field": "metadata.doc_type", "operator": "eq", "value": "contract"},
]},
pagination={"method": "cursor", "limit": 10},
)
for doc in results["documents"]:
print(doc["document_id"], doc.get("page_number"), doc.get("object_type"), doc["score"])Questions about Mixpeek for document intelligence
What document formats does Mixpeek support for intelligent processing?
Mixpeek processes all common document formats including PDF (native and scanned), Microsoft Word (DOC, DOCX), Excel spreadsheets (XLS, XLSX), PowerPoint presentations (PPT, PPTX), plain text, HTML, and image files containing text (JPG, PNG, TIFF). Scanned documents are processed through enhanced OCR that handles skewed pages, handwriting, stamps, and low-resolution scans. Documents can be ingested from S3, GCS, Azure Blob Storage, or via direct API upload.
How does Mixpeek handle complex document layouts like tables and multi-column formats?
Mixpeek uses layout-aware extraction models that detect and preserve document structure including tables, columns, headers, footers, sidebars, and nested lists. Table extraction captures row and column relationships, merged cells, and header mappings as structured JSON. Multi-column layouts are correctly segmented so text flows in reading order rather than being merged across columns.
Can I build custom document classification taxonomies?
Yes. Mixpeek supports custom taxonomy creation through labeled training data or zero-shot classification using natural language descriptions of your categories. Common enterprise taxonomies include document type (contract, invoice, memo, report), sensitivity level (public, internal, confidential, restricted), regulatory category (GDPR, HIPAA, SOX), and department-specific classifications. Custom taxonomies can be applied alongside standard classifications.
How does cross-document semantic search differ from traditional full-text search?
Traditional full-text search matches keywords and requires exact or fuzzy string matches. Mixpeek semantic search understands meaning, so a query for 'termination clauses with 30-day notice' finds relevant paragraphs even when the document uses different wording like 'cancellation provisions requiring one month advance written notice.' Semantic search also works across document types, finding related content in contracts, memos, and email attachments simultaneously.
What is the extraction accuracy for structured fields like dates, amounts, and party names?
For well-formatted digital documents, field extraction accuracy exceeds 99% for common fields including dates, monetary amounts, company names, addresses, and reference numbers. Scanned documents achieve 95-98% accuracy depending on scan quality. All extractions include confidence scores, allowing you to route low-confidence results to human review while auto-processing high-confidence extractions.
How does Mixpeek handle document versioning and change detection?
Mixpeek can process multiple versions of the same document and identify differences at the paragraph, clause, and field level. This is particularly valuable for contract redlining, policy update tracking, and regulatory filing comparisons. Change detection works across formats, so you can compare a Word document against a scanned PDF of an earlier version.
Can Mixpeek extract data from handwritten documents or annotations?
Yes. Mixpeek includes handwriting recognition models that process handwritten notes, annotations, signatures, and form fields. Accuracy varies by handwriting legibility, but the system handles common use cases including handwritten form entries, margin notes on printed documents, and signed agreement pages. Confidence scores flag low-legibility content for human review.
How does entity extraction and relationship mapping work across documents?
Mixpeek extracts named entities including people, organizations, locations, dates, and monetary amounts from every processed document. Entities are deduplicated and linked across the corpus, building a relationship graph that reveals connections between parties, agreements, and events that span multiple documents. This is especially valuable for due diligence, litigation support, and compliance investigations.
What security and compliance certifications does Mixpeek hold for document processing?
Mixpeek holds GDPR compliance. It holds neither SOC 2 nor HIPAA certification and has no audit under way; the current status of each is published at mixpeek.com/trust. Document data is encrypted in transit (TLS 1.3) and at rest (AES-256). The hosted service processes in the United States, self-hosting covers stricter residency requirements, access is role-based, and audit logs track every document access and processing event. On-premise deployment is available where requirements are stricter.
How does pricing work for document intelligence?
Pricing is based on document volume and processing features enabled. Basic extraction and classification starts at lower tiers suitable for teams processing hundreds of documents monthly. Enterprise plans support millions of documents with dedicated infrastructure, custom model training, and priority support. All plans include semantic search across your processed corpus. Contact us for a custom quote based on your volume and requirements.
Build this in the docs
The extractors and retrieval stages this solution runs on, with API reference and worked examples.
Run this on your own enterprise & legal content
Point Mixpeek at the storage you already have and search your video, images, audio and documents the way this page describes.