Antfly as an inference pipeline.

Flip the reranker on

Keyword and semantic search each rank the same articles differently, and fusion merges the two lists. A reranker then reads the candidates themselves and scores each one against the question. The switch swaps in the reranked list from the same run.

Pull out the entities

Antfly can also read each document as it arrives and pull out the things it names: people, places, universities, theories. You say which kinds matter to you, and a small extraction model finds them, with no rules to write. Below is the result over sixteen more Wikipedia articles, the introductions of physicists from Newton to Hawking. Eight of them share the Nobel Prize node, four share Cambridge, and where one physicist’s introduction names another, the two link directly: Rutherford to Bohr, Bohr to Heisenberg, Pauli to Einstein and Fermi. Store those links in a graph index and a query can walk from one article to the others through what they mention.

Inside the engine

A retrieval pipeline usually grows one piece at a time: a chunker, an embedding API, a vector index, a keyword engine beside it, and a reranker, each with its own copy of the document ids. In Antfly they are stages of one engine over one set of ids. You keep your sources and your generative model; the rest runs as documents arrive.

Chunking

Text and multimodal content split for embedding.

Embedding

Local models for text, images, and audio; no per-request API.

Extraction

An extraction model (GLiNER2) pulls the people, places, and dates a document names into typed fields you can filter and aggregate on.

Indexing

Full-text, vector, graph, and typed values on the same documents.

Fusion

Keyword and vector results merged by rank or by score.

Reranking

A cross-encoder scores the candidates in one batched call.

Pruning

The weak tail is dropped before the result leaves the engine.

Change the model, keep the table

Chunks, embeddings, transcripts, and extracted entities are artifacts derived from the source document, and any index can build from them. Swap the embedder and the engine re-embeds from the stored chunks; you never re-insert a document to change the embedder. Each index reports its own readiness, and a document whose enrichment failed is listed as failed until it is reprocessed.

Inputs

Text and JSON

Inserted directly or by URL.

PDFs

Born-digital parsing, with OCR for pages that carry no text.

Images

CLIP-style embeddings, so a text query can return an image.

Audio

Whisper transcribes it, and the transcript is searched as text.

• Tune hybrid searchFusion, reranking, and pruning, added one stage at a time
• Build artifact-backed indexesHow chunks, embeddings, and relationships derive from the source document
• Browse modelsEvery embedder, reranker, and extractor Antfly Inference runs