jammi-ai 0.16.0

Embeddable AI engine for inference, embeddings, vector search, and fine-tuning

Jammi AI

Jammi is an embeddable AI engine that brings model inference into your data pipeline. Register data sources, run SQL queries, generate embeddings, search with vector similarity, fine-tune models on your domain, and evaluate results — all without leaving your application.

Install

pip install jammi-ai

For CUDA/GPU support:

pip install jammi-ai-cu12

Quickstart

The 5-minute walkthrough — install, connect, register a source, generate embeddings, search — lives in cookbook/quickstart/ with a runnable quickstart.py gated by CI. The condensed version:

import jammi_ai

db = jammi_ai.connect(gpu_device=-1)
db.add_source("corpus", url="cookbook/fixtures/tiny_corpus.parquet", format="parquet")

MODEL = "sentence-transformers/all-MiniLM-L6-v2"
db.generate_text_embeddings(source="corpus", model=MODEL, columns=["content"], key="id")

query_vec = db.encode_text_query(MODEL, "quantum computing applications")
results = db.search("corpus", query=query_vec, k=5).run()
print(results.to_pandas())

For runnable end-to-end recipes — mutable tables, trigger streams, eval, fine-tuning, Flight SQL — see cookbook/.

Features

• SQL over local files — query Parquet, CSV, and JSON via DataFusion
• Federated queries — join local files with PostgreSQL or MySQL
• Text embeddings — load any BERT-family model from Hugging Face Hub (or local safetensors / ONNX) and persist results to Parquet with ANN indexes
• Image embeddings — CLIP-style vision encoders
• Vector search — ANN similarity search with automatic brute-force fallback
• SearchBuilder — fluent API for .filter(), .sort(), .join(), .annotate(), .limit(), .select(), .run()
• Evidence provenance — retrieved_by and annotated_by tracking on every search result
• Fine-tuning — LoRA / deep LoRA adapters with contrastive loss to improve embeddings for your domain
• Evaluation — recall@k, precision@k, MRR, nDCG, accuracy, F1, and A/B model comparison
• Model caching — LRU eviction, ref-counted guards, single-flight loading
• GPU scheduling — memory-budget admission control with RAII permits
• Crash recovery — recovers embedding tables stuck in "building" state on restart

SearchBuilder

search = db.search("patents", query=query_vec, k=20)
search.filter("year >= 2020")
search.sort("similarity", descending=True)
search.limit(5)
search.select(["id", "title", "similarity"])
results = search.run()   # pyarrow.Table

All results are returned as pyarrow.Table — zero-copy from the Rust engine.

Fine-tuning

job = db.fine_tune(
    source="patents",
    model="sentence-transformers/all-MiniLM-L6-v2",
    triplets="triplets_train.parquet",
)
job.wait()

Requirements

• Python 3.9+
• Linux (x86_64) or macOS (Apple Silicon or Intel)

Windows is not yet supported due to a dependency on POSIX memory-mapping APIs.

Running the OSS server

For deployments that need a long-running Flight SQL + gRPC service rather than an embedded library, the workspace ships a Docker image:

docker run --rm \
  -p 8080:8080 -p 8081:8081 \
  -v jammi_data:/var/lib/jammi \
  ghcr.io/f-inverse/jammi-ai-server:latest

curl http://localhost:8080/healthz
# {"status":"ok","version":"0.8.0"}

The OSS server is single-tenant — the deployer's network is the auth boundary. See Deploy as a Server for the full guide.

Documentation

Full documentation, including guides for SQL queries, embeddings, search, fine-tuning, and evaluation:

https://f-inverse.github.io/jammi-ai/

For the engine's design philosophy — what belongs in Jammi versus a consumer's own repo, how embeddings are consumed, and how it deploys — see Design Philosophy.

License

Apache-2.0