| --- |
| language: |
| - en |
| license: apache-2.0 |
| library_name: sentence-transformers |
| pipeline_tag: feature-extraction |
| base_model: |
| - BAAI/bge-m3 |
| datasets: |
| - Cilow/Vinci-Evals |
| tags: |
| - sentence-transformers |
| - text-embeddings-inference |
| - embeddings |
| - retrieval |
| - matryoshka |
| - lattice |
| - cilow |
| - db-native |
| - claim-aware |
| --- |
| |
| # Vinci |
|
|
| Vinci is Cilow's DB-native, claim-aware embedding model family for the Lattice |
| retrieval pipeline. This repo is the canonical Hub target for `Cilow/Vinci`. |
|
|
| Current status: **experiment-runner scaffold**. Do not treat this repo as production-ready until |
| the shadow evaluation report shows no regression against the current production |
| baseline and the repo contains trained model weights. |
|
|
| ## Model Contract |
|
|
| - Model id: `Cilow/Vinci` |
| - Legacy aliases: `cilow/vinci-v0, cilow/cilow-embed-v0` |
| - Base model: `BAAI/bge-m3` |
| - Full dimension: `1024` |
| - Low-dimension Matryoshka prefix: `256` |
| - Similarity: cosine |
| - Serializer: `claim_v1` |
| - Pipeline: `Lattice` |
| - Fresh index required when switching semantic spaces: yes |
|
|
| ## Intended Use |
|
|
| Vinci is designed for retrieval over structured application context: factual |
| claims, memories, code snippets, tool traces, database rows, temporal updates, |
| provenance, confidence, and supersession state. It is not a general-purpose |
| similarity model by default; promotion depends on Cilow/Lattice retrieval gates. |
|
|
| ## Training Plan |
|
|
| The first training run fine-tunes `BAAI/bge-m3` with |
| `SentenceTransformerTrainer`, `MultipleNegativesRankingLoss`, and |
| `MatryoshkaLoss([1024, 256])`. The dataset policy is synthetic plus explicit |
| eval fixtures only. Private user memories are excluded. |
|
|
| ## Data Policy |
|
|
| - Private user data: `False` |
| - Allowed sources: synthetic and checked evaluation fixtures |
| - LongMemEval-like data: eval-only unless an explicit train/dev/test split is approved |
| - Hard negatives: superseded facts, wrong predicate/entity, numeric near misses, |
| temporal stale/current conflicts, and semantic neighbors with factual mismatch |
|
|
| ## CPU Smoke Metrics |
|
|
| | Backend | Dim | Recall@5 | Recall@10 | MRR@10 | nDCG@10 | |
| |---|---:|---:|---:|---:|---:| |
| | deterministic | 256 | 0.1429 | 0.7143 | 0.1160 | 0.2484 | |
| | deterministic | 1024 | 0.2857 | 0.7143 | 0.1262 | 0.2581 | |
|
|
| These CPU metrics use the deterministic backend unless otherwise stated. They |
| are a pipeline smoke test, not a production quality claim. |
|
|
| ## Dataset Splits |
|
|
| | Split | Queries | Triplets | |
| |---|---:|---:| |
| | test | 2 | 12 | |
| | train | 5 | 30 | |
|
|
| ## Deployment |
|
|
| The preferred shadow path is TEI with `TEI_MODEL=Cilow/Vinci` and |
| `TEI_DIMENSION=1024`. Local inference uses the ONNX lane only after export and |
| dimension validation pass. Lattice must encode model id, serializer version, |
| dimension, corpus version, and index version in metadata to prevent mixed-vector |
| comparisons. |
|
|
| ## Limitations |
|
|
| - No private data is included in the starter dataset. |
| - A trained checkpoint must still pass local SentenceTransformer eval, TEI eval, |
| and a production-shadow comparison before promotion. |
| - Voyage or the current production provider remains default until Vinci matches |
| or beats the baseline without latency regression. |
|
|
| Generated: `2026-05-02T02:51:25+00:00` |
|
|
