SPARK Relay
A Software Spark-developed retrieval model line for enterprise search, evidence selection, and governed retrieval pipelines.
VERSION0.8FAMILYEmbedding and rerankingCONTEXTUp to 32K target; confirmed per deploymentRELEASEControlled preview
- INPUTS
- Text queries, passages, document metadata
- OUTPUTS
- Embeddings, relevance scores, ranked evidence
- OWNERSHIP
- Software Spark-developed model program; artifact access subject to evaluation agreement.
- Enterprise semantic search
- Retrieval-augmented generation
- Evidence selection and reranking
- Access-aware knowledge workflows
- Retrieval quality depends on corpus and chunking design
- Relevance does not guarantee factual correctness
- Entitlements must be enforced outside the model score
- Domain adaptation requires representative evaluation data
- Recall@K and nDCG on disclosed datasets
- Evidence coverage for downstream answers
- Entitlement-filtering integration tests
- Latency and index-size scaling
No external benchmark score is claimed on this card. Reproducible results will follow the SPARK benchmark protocol.