How it works

Your data, without the blind spots.

Write a recording, a scan, a slide deck, or plain text. The models inside Antfly transcribe, read, extract, and embed it as it arrives, and one query searches everything they found.

Input

AntflyAntfly

Inference Engine

  • Transcribe, with speakerswhisper-small
  • Read pagesLightOnOCR-3
  • Chunk to the model's limitsmodel-aware
  • Extract entities and relationsgliner2.5
  • Embed text and imagesembeddinggemma-2

Database

  • Table
  • Full-text index
  • Vector index
  • Knowledge graph
Models on the indexes
You name the models once, in a table's indexes: OCR, transcription, extraction, chunking, embedding, each sized to its job. After each write Antfly runs them in the background and keeps what they produce with the data. Apps, agents, and queries read that output without running a model. To try a different model, add an index that names it: it fills from the documents already stored, and you drop the old one when you're done.
One engine for every index
Documents, relational tables with joins, graph, and key-value in one engine, with full-text and vector search over all of it. What the models extract lands in every one of them: text read from a scan is searchable, and fields pulled from a contract join with rows.
Open source
Apache-2.0, all of it, built in public on GitHub. You and your agents can read the code, find what broke, and fix it.

What it replaces

Without it, each step is its own system: a transcription service, an OCR service, an embedding API, a vector store, a search engine, a graph database, and the database with your records, plus the code that moves data between them.

Without Antflythe models run in pipelines outside the databases, and each kind of question lives in its own system, kept in sync by code.

With Antflyone engine runs the models and holds every data model.

Each of those systems was a reasonable choice. The cost shows up later: a new question needs new structure first, every consumer rebuilds the pipeline it needs, and a question that spans them gets stitched together in application code.

Automatic knowledge graphs

Antfly can build a knowledge graph for you. The models extract the entities and relations in each document, resolve mentions across documents into one record per entity, and keep the edges current as documents change. Nobody supplies the edges, and each entity record is an ordinary document a person can correct.

The graph sits beside everything else: traverse it, search it by full-text and vector, and join the extracted fields with rows. The pipeline page shows one the models built from Wikipedia articles, and the knowledge graph guide builds one from your documents.

See it answer a question

Bigger questions walks through real questions end to end: the file, what the models make of it, and every table, index, and access rule behind the answer.

Prototype to production

Start on one node.

Install, run one command, and query in minutes. When you outgrow the node, your code comes with you.

Run your first query