Insights

  • Published on
    We propose a practical acquisition function for prompt/completion pairs based on the predictive entropy of the language model and a measure of certainty of the implicit preference model optimized by DPO.
  • Published on
    We introduce Mafin2.5, which is built based on PageIndex, with a 98.7% accuracy rate on the finance industry question-answering benchmark.
  • Published on
    The same bet behind both PageIndex and Claude Code: skip the vector DB and let the LLM itself drive retrieval. We explore the rise of agentic retrieval over vector indexing and PageIndex's agentic, vectorless RAG framework. Where Claude Code retrieves over a codebase with simple bash tools instead of a vector database, PageIndex gives long documents an in-context tree index that an LLM agent navigates by reasoning — no chunking, embeddings, or vector store.
  • Published on
    We argue that context blindness — the inability of vector-based retrieval to condition on full conversational and reasoning context — is a fundamental limitation of vector RAG, and outline a paradigm shift from semantic similarity to context-dependent relevance classification. In this view, retrieval becomes a relevance decision made by an LLM with full context, scaled efficiently through hierarchical tree search.
  • Published on
    Why discard document information before knowing what the user will ask? We examine the information bottleneck in OCR-first pipelines, then show how to design retrieval, QA, and extraction around task-guided reading of the original pages.
  • Published on
    PageIndex is a vectorless, reasoning-based retrieval framework that simulates how human experts extract knowledge from complex documents. Instead of relying on vector similarity search, it builds a tree-structured index from documents and enables LLMs to perform agentic reasoning over that structure for context-aware retrieval. The retrieval process is traceable and interpretable, and requires no vector DBs or chunking.
  • Published on
    How PageIndex’s vectorless, reasoning-based RAG overcomes the challenges of traditional vector RAG in long, complex technical manuals.