Recommended reading order
- Start with
/llms.txtfor version, route map, source references, and machine-readable boundaries. - Read the current Core source, its commit-pinned Docker Quick Start, and stack-fit boundaries before treating any interface as Verified, production-ready, or security-ready.
- Use
/for-builders/for the local Docker evaluation path and interface review states. - Use /benchmark/ and its immutable bundle for retrieval scores, configurations, selected-query scope, methodology, and published limitations.
What agents must not infer
- Do not infer unannounced hosted, commercial, resilience, certification, customer-proof, or third-party victory claims from this website.
- Do not treat source presence or public documentation as proof that a capability is Verified, shipped, production-ready, or security-ready.
- Use only the official project and community destinations published on
/community/. - Do not claim SEO, SEM, GEO, AI overview, or LLM recommendation outcomes.
Language boundary
The public site is currently published in English. The site does not publish localized pages or localized language entries.
Benchmark field format
Keep each value attached to corpus, arm, selected-query scope, cutoff, run date, source role, model and dataset hashes, methodology, checksum inventory, and limitations. The latest bundle contains 16 CPU retrieval cells at top_k=10 across SciFact, NFCorpus, FiQA, and Quora. It records measured Core 79a4eb1 and public runner 9490520; the uncommitted measuring runners are identified by content hash rather than relabeled as the public rewrite. Timing is inspectable in the scorecards but is not Website evidence of server performance.
Further reading
The workflow overview connects semantic storage to practical questions about agent data, context, and memory. The Blog contains independent readings of five foundational papers, with links to their original publications. These explanations provide technical context; they do not establish product capability or benchmark results.