Evaluate one retrieval path
Vector, keyword, fusion, and reranking paths are present in source. Use the full-corpus evidence for retrieval quality and verify runtime behavior at the pinned commit.
Inspect records without assuming isolation
Memory APIs and typed-record paths are present in source. Fact extraction and security isolation are not currently verified capabilities.
Inspect links back to sources
Source-context and trace surfaces are present for evaluation. Treat end-to-end attribution and production readiness as review required.
Builder paths
The current routes separate a documented evaluation path from interfaces that remain under review or Blocked.
Docker Quick Start
Follow Quick Start to build the checked-out source, download pinned BGE-M3 embedding and bge-reranker-v2-m3 reranking assets, wait for source-backed readiness, and send a query with rerank=true. The path publishes only 127.0.0.1:8420; external LLM roles and the built-in Agent remain disabled. It is not a production-readiness, parser-coverage, or retrieval-quality benchmark.
Existing stack
Keep operational systems in place and review the commit-pinned stack-fit decision cards for data movement, sync ownership, failure domains, security responsibility, and rollback boundaries. They do not claim universal compatibility or replacement.
Retrieval evidence
Open the semantic storage retrieval benchmarks for commit-pinned retrieval scores, configurations, selected-query scope, methodology, and retrieval-only boundaries.
Understand context and memory before integrating
Start with the semantic storage overview. For the underlying mechanisms, read how RAG uses external context and how Reflexion retains feedback as agent memory.