Evoke-Model-Beta-1

Evoke Model (Beta 1)

This repository distributes the frozen P2.2 ABI-v2 model checkout used by Evoke, a PostgreSQL-native BM25 engine with optional Sparse Semantic Retrieval (SSR) postings in the same index.

This is the existing release model, published for reproducible source and package builds. It is not a newly trained checkpoint, a standalone database, or a generic Sentence Transformers model directory. Use the Evoke runtime and its tokenizer, calibration, lexical vocabulary, and scoring contract together.

Model And Runtime

The sparse route derives from IBM's approximately 30M-parameter Granite sparse encoder. Evoke supplies separate document/query ONNX exports and combines semantic atoms with exact lexical atoms in one sparse posting namespace. Document completion can run in background workers; queries use the same frozen model identity as the index.

The _smoke suffix is part of the frozen model ID, not a placeholder to edit. Changing model artifacts or their contract requires a separately qualified checkout and index rebuild. Downloading this identical checkout does not by itself require rebuilding existing indexes.

Files

checkout/
  manifest.json
  encoder/
    semantic_document_compiler.onnx
    semantic_query_compiler.onnx
    semantic_runtime.json
    tokenizer.json
    tokenizer_config.json
    merges.txt
    vocab.json
  lexical/vocabulary.json
  sae/atom_space.json
  calibration/
    query_calibration_rms.f32
    query_calibration_stats.npz
    scoring_profile.json
  provenance/p2_contract.json
evoke-p2.2-nfcorpus-v2.zip
evoke-p2.2-nfcorpus-v2.zip.sha256
milestone-model.json
LICENSE
NOTICE

The ZIP contains exactly the same checkout under evoke-p2.2-nfcorpus-v2/. The separate file tree makes the ONNX graphs and metadata inspectable; the ZIP supports the existing Evoke package/CI downloader without introducing a new loader or changing the model contract. milestone-model.json is the source repository's content lock.

Model-card, license, Hub cache, and Git metadata are intentionally outside checkout/: the release validator requires an exact model file inventory. Frozen provenance labels record the source audit fixture identity; they are not runtime dependencies or additional required downloads.

Download And Build

Public downloads do not require a Hugging Face account or access token. From an Evoke source checkout, install the official hf CLI if needed and run:

hf download Intelligent-Internet/Evoke-Model-Beta-1 \
    evoke-p2.2-nfcorpus-v2.zip evoke-p2.2-nfcorpus-v2.zip.sha256 \
    --local-dir .artifacts/evoke-model-download

For repeatable builds, add --revision with the full Hub commit SHA from the repository's Files and versions page. Use the ZIP, not the download directory containing Hub cache metadata, as the input to the Evoke fetcher:

python3 scripts/fetch_milestone_model.py \
    --url "file://$PWD/.artifacts/evoke-model-download/evoke-p2.2-nfcorpus-v2.zip" \
    --archive-sha256 "$(awk '{print $1}' .artifacts/evoke-model-download/evoke-p2.2-nfcorpus-v2.zip.sha256)" \
    --output .artifacts/evoke-milestone-model
python3 scripts/validate_milestone_model_checkout.py \
    --checkout .artifacts/evoke-milestone-model

The fetcher checks the archive checksum and the source-controlled manifest, all artifact digests, and exact inventory before installing the checkout. It refuses to overwrite an existing output by default. See Evoke's contributor guide for ZIP/Docker builds and the pinned ONNX Runtime SDK.

Compiling the extension does not require these model weights. Complete release ZIPs and Docker images do; source-installed SSR indexes also need an administrator-installed checkout. Ordinary exact BM25 indexes do not perform model inference.

Evaluation Boundary And Direction

P2.2 is the full-text, lifecycle-qualified package successor to P2.1. The project's historical P2.1 BEIR15/MTEB10 comparisons are not measurements of this P2.2 checkout under every corpus, storage profile, or query workload. See the project's technical reports through its documentation index for the evidence and reproduction boundaries.

The upstream encoder is English-focused. This checkout's lexical vocabulary and query calibration are frozen from NFCorpus; domain transfer and other languages require separate evaluation. The upstream model card documents both public and non-public training sources; this release distributes the derived runtime artifacts, not the upstream training corpus. Future work focuses on broader corpus calibration, transfer quality, and the cost of high-document-frequency semantic postings.

License And Attribution

Distributed under Apache License 2.0. The upstream model was developed by IBM's Granite Embedding Team. See NOTICE for the exact upstream revision and Evoke modifications, and the upstream model card for its attribution, intended uses, and training-data description.