Parsers, NER models and vector DBs are generic. indx builds your context layer.
Built for Japanese business documents, and run in your own environment. You define the document types and fields. indx sends each page through the cheapest method that reads it, keeping OCR and vision models for the pages that need them, and every correction your team makes becomes a label the next router version is scored on.
From your documents to your context layer, and back
The blueprint's diagram, each part coloured by what answers today.
Three parts, one loop
Enrich your data, shape your embeddings, build a router for each job and try variants.
One document at a time: how it was parsed, what was extracted, how it was tagged and summarized, and what it matches.
Every document in aggregate: accuracy by type and task, volume, the review queue and versions.
Why not a parser, an NER model or a vector DB?
Categories, not products, and no benchmark.| Document parser / IDP | NER / extraction model | RAG / vector DB | indx | |
|---|---|---|---|---|
| Japanese business documents | Generic layouts | Per-language model | Model default | DoneJapanese taxonomies and entity lanes, 工程系統図, Shift_JIS; benchmarked on Japanese filings and notices |
| OCR and vision models only where needed | Every page, one pipeline | — | — | DoneThe ladder reads a text layer on CPU first; OCR, a vision model, then a person only below it |
| Your own document types and fields | Vendor catalogue | Model labels | — | WIPTypes, fields, reviewed labels and datasets on Data & labels; labels at scale are phase 9 |
| A different method per document type | Fixed pipeline | One model | One chunker | DoneA router gives each type its own reader, chunkers, lanes and spaces |
| See the route before it runs | — | — | — | DonePOST /v1/plan, per page |
| Define what “similar” means | — | — | Model default | PlannedFour shipped spaces today; your own is phase 6 |
| Compare methods on your own data | — | — | — | WIPExperiments score candidates on your evaluation set and promote one; the assistant is phase 10 |
| Improves with every correction | Vendor releases | Retraining project | — | WIPCorrections become labels and experiments promote the better route today; learning from them on its own is phase 5 |
| Runs in your environment | Usually SaaS | Varies | Varies | DoneOne image with its weights baked in, deployed on AWS, Azure and Google Cloud, and runnable on your own servers |
| Replace any part | Vendor's engines | One model | Varies | DoneEvery reader, chunker, classifier and store is a plugin, over open-source engines |
indx does not replace these tools: it runs them as methods and chooses between them per page.