This baseline model generates responses using only its internal weights. It does not have access to external documents or real-time data, relying entirely on its pre-trained knowledge base.

Ideal for general reasoning tasks where external context is not required. Pick a model above and hit SUBMIT to try it for real.

A standard RAG implementation that retrieves relevant document chunks based on vector similarity and appends them to the prompt context before generation.

Provides factual grounding but may struggle with complex multi-hop queries or large document sets. Pick a model above and hit SUBMIT to try it for real (first request loads the embedding model, ~20s).

Chunking + hybrid (sparse + dense) retrieval + cross-encoder reranking, top_k=10. Optimized for high precision by re-scoring a wider candidate pool before generation.

Pick a model above and hit SUBMIT to try it for real (first request loads 3 models: chunk encoder, sparse index, reranker — ~60s; reranking itself also takes ~30-50s per query on CPU).

LLM decomposes the question into sub-issues, searches + reranks each one independently, then keeps only the global top-3 highest-reranked passages across all sub-issues for generation.

Pick a model above and hit SUBMIT to try it for real (shares models with Advanced RAG; each sub-issue adds another rerank pass, so this is the slowest card — often 60-120s+/query).