The problem
Research corpora do not fit in a context window. Stuffing a million tokens into one prompt fails on length; summarizing first loses the details a researcher actually needs; and either way, the reasoning happens invisibly inside one model call that can't be audited. PRISM — Process-Recursive Inference with Step-level Monitoring — takes a different route: decompose the reasoning into code, and trace every step.
Architecture
PRISM is a production-grade implementation of the Recursive Language Model paradigm (Zhang, Kraska, Khattab — MIT, Dec 2025). Claude Sonnet acts as the root model, but instead of answering directly it writes executable Python in a REPL. That code decomposes the query, dispatches 10–50 sub-LM calls per query to Qwen3-8B — served locally via vLLM with PagedAttention, so sub-calls cost nothing — and synthesizes the results. Iterative chunking and synthesis let the pipeline work through corpora past 1M tokens, far beyond any single context window.
Retrieval is hybrid: Qdrant stores dense (BGE) and sparse (SPLADE) embeddings, fused with Reciprocal Rank Fusion. Ingestion runs arXiv PDFs through Unstructured.io into the index; a Gradio frontend sits on top.
What's measured
Every REPL iteration, sub-LM call, and code execution is logged to Langfuse — 100% step tracing, so any answer can be replayed down to the exact sub-calls that produced it. Evaluation uses RAGAS (faithfulness, answer relevancy, factual correctness) with a benchmark runner and a GPT-4o baseline comparison script, keeping the recursive system honest against the obvious single-model alternative.
Honest limitations
The repo ships the benchmark harness but does not publish headline scores yet, so no accuracy number is quoted here — the RAGAS comparison is reproducible from the repo, not summarized on this page. Phase 2 — post-training Qwen3-8B with GRPO so recursive decomposition becomes a native behavior instead of root-model-driven orchestration — is planned, not built. And the zero-cost sub-LM claim is about API spend, not hardware: vLLM serving Qwen3-8B requires a local GPU.