Skip to content

Guides

Practical guides for operating Uni-Xervo in production environments.

  • Provider Selection: choose providers by capability, latency, and control.
  • ONNX Runtime: use local/onnx for raw tensor execution, HF snapshots, and ONNX Runtime session management.
  • Config Validation: enforce schema correctness in CI and startup.
  • Multimodal Generation: vision, diffusion, and speech pipelines with local/mistralrs.
  • Structured NLP: POS / NER / dependency / SRL / dialog-act analysis via NlpModel and the kniv-deberta cascade.
  • Multimodal Trait Surface: the full task-trait surface — the ModelInfo supertrait plus the image / audio / multimodal / sparse / multi-vector embed, NLP, document extract, transcription, and OCR traits.
  • Sparse Embeddings: learned-sparse term-weight vectors (SPLADE / BGE-M3 sparse) via SparseEmbeddingModel, for the lexical half of hybrid retrieval.
  • Multi-Vector Embeddings: per-token (ColBERT / late-interaction) vectors via MultiVectorEmbeddingModel, scored with host-side MaxSim.
  • OCR: optical character recognition with bounding boxes and confidences via OcrModel and the local/onnx PP-OCR pipeline.
  • Tiered PDF Extraction: the uni-xervo-pdf companion crate — escalate per page across native text → OCR → doc-VLM, with provenance and cross-tier verification.

For full ONNX developer documentation, use the dedicated ONNX section.