01
No sidecar service
The model runs inside the DB process. VECTOR_EMBEDDING(MODEL USING :text AS DATA) is a function call — no HTTP, no queue, no fanout.
v0.2 — Oracle AI Database
Embeddings & rerankers that live inside your database.
onnx2oracle packages any HuggingFace sentence-embedding model — and now cross-encoder rerankers —
as an augmented ONNX graph and loads it directly into Oracle via DBMS_VECTOR.LOAD_ONNX_MODEL.
Embeddings via VECTOR_EMBEDDING, relevance scoring via PREDICTION. No sidecar, no queue, no network hop.
Why
In-database embeddings, for one reason at a time.
Most vector pipelines bolt an embedding microservice in front of the database. That adds a second system to deploy, secure, patch, and observe. Oracle AI Database can run the model itself — if you can get it in.
02
Data never leaves the tablespace
Compliance teams like this. Your document column gets embedded without a byte crossing a network boundary.
03
Transactional consistency
Insert a row, embed it, index it — all inside the same transaction. No dual-writes, no eventual sync drift.
04
One binary, seven models, zero YAML
onnx2oracle presets shows the shelf — 5 embedding models plus 2 reranker presets. load <name> does the rest, including tokenizer wrapping, L2 normalization (embeddings), or [CLS] q [SEP] d [SEP] splicing (rerankers).
Presets
Seven vetted models. Embed, or rerank, or both.
Five embedding models plus two cross-encoder rerankers. Each preset maps a HuggingFace repo to an Oracle mining-model name. Dimensions, pooling, and task type match the model card — not a rewrite.
all-MiniLM-L6-v2384-d
The default. Small, fast, good enough for English-only semantic search up to a few million rows.
all-MiniLM-L12-v2384-d
Same geometry as L6, deeper transformer. Better recall on longer documents; ~30% slower.
all-mpnet-base-v2768-d
The quality pick for English. Double the vector width, heavier index. Worth it when MRR matters.
bge-small-en-v1.5384-d
Strong on BEIR benchmarks. Uses CLS pooling, not mean — the loader handles the difference.
nomic-embed-text-v1768-d
Long-context (8192 tokens) with Matryoshka-friendly output. Biggest curated preset — plan the disk.
ms-marco-MiniLM-L-6-v2reranker
Cross-encoder reranker. Loaded with function:"regression"; queried with PREDICTION(model USING q AS DATA1, d AS DATA2) for two-stage retrieval.
ms-marco-MiniLM-L-12-v2reranker
Deeper cross-encoder reranker. Higher quality at ~30% more latency than L-6. Same WordPiece tokenizer, same SQL surface.
The guide
Five short pages, in order.
Each page assumes the previous one worked. If something breaks mid-way, the troubleshooting reference maps ORA-codes to fixes.
01
Quickstart
Docker up, load the default model, verify the round trip in about four minutes.
02How it works
The augmented ONNX pipeline — tokenizer, pool, L2-norm — explained with a diagram.
03Pick a model
Decision tree across embedding + reranker presets: English vs. multilingual, 384 vs. 768, speed vs. recall, embed vs. rerank.
04Oracle ADB wallet flow
Point the loader at Autonomous Database. Walletless mTLS is shortest; classic wallet still works.
05Custom models
The --from-huggingface escape hatch with worked examples, including CLS-pooled models.
CLI reference
Every flag, every subcommand, including DSN resolution precedence.