Shipped

PageIndexOllama - Local-First Fork of PageIndex

Open Source

⚡All inference required a live OpenAI API key - offline or air-gapped execution was blocked entirely; provider switches corrupted traversal silently with no error surface; token encoding differences across providers produced inconsistent chunk boundaries with no visible signal.

⚠️PageIndex's tree RAG was hardcoded to OpenAI's API contract - inline prompt strings, non-normalized completion handling. Local or offline deployment was impossible. Any provider change broke traversal.

⚙️Forked and refactored: provider-routing abstraction resolved via env vars. Finish-reason normalization layer stabilizes recursive traversal across model outputs. Prompts externalized into a registry loader. Bounded async concurrency across TOC generation and summarization. Hierarchical fallback for large-document robustness.

🛡️Normalized completion contracts prevent finish-reason variations from corrupting recursive traversal state. Fallback chunk policies handle constrained VRAM and RAM. Structured-output hardening absorbs imperfect model responses without pipeline failure.

🚀Fully offline tree-RAG with Ollama - no API keys. Seamless provider switching via stable internal contracts. Regression risk reduced through e2e coverage across document types and model sizes.

↳ Fully offline tree-RAG execution • vendor lock-in eliminated • provider-agnostic runtime

Provider AbstractionTree-Based RAGOllama / Local LLMsBounded Async ConcurrencyHierarchical FallbackPython / Open Source

View on GitHub