Multimodal retrieval engine
Retrieve. Rerank. Ground.
Mosaic assembles text and image chunks into cited, grounded answers — with pluggable embeddings, hybrid retrieval, rerankers, and honest benchmark ablations.
Pluggable pipeline
Swap embedders, retrievers, rerankers, and chunkers independently.
Hybrid retrieval
Dense + BM25 fusion for keyword-heavy enterprise documents.
Rigorous evals
Recall@k, MRR, faithfulness, latency, and cost per query.
Built like internal FAANG tooling
Typed Python API, pytest coverage, ruff + mypy, CI on every push, and reproducible one-command benchmarks — not a notebook demo.
pip install mosaic-rag mosaic-benchmark --output benchmarks/results