Pablo Marcos Parra

Blog

Notes from production.

Write-ups on LLM systems, agents and the unglamorous work that makes them reliable.

  1. One passage, one yes/no question

    I put two decision models, Jev and Laya, up against the LLM reranker of my D&D wiki. Jev, reading full passages, found the right source for all 36 questions. Laya without fine-tuning did worse than not reranking at all.

    12 min read
    • rag
    • reranking
    • llm-local
    • ollama

  2. Retrieval isn't the problem. Ranking is.

    How retrieval works in the Oracle of my D&D wiki: three search paths fused by rank, a reranker that only helps as long as it never sees the raw text, and leftover slots for what didn't fit.

    12 min read
    • rag
    • llm-local
    • ollama
    • chromadb
    • reranking

  3. Keeping the harness, swapping the model

    What it takes to run Claude Code against NVIDIA NIM and local Ollama models when the subscription quota runs out and why the model that benchmarks best is not the one that works.

    8 min read
    • claude-code
    • litellm
    • local-llm
    • tooling