gos-skill-retrieval-embeddings

Graph of Skills retrieval vectors

Fixed text-embedding-3-large vectors (3,072 dimensions, float32) for the skill libraries and public dev queries used by the OpenRSI task rsi/gos-skill-retrieval. They let Graph of Skills default retrieval run offline: the task's starting retriever seeds personalized PageRank from these vectors and lexical candidates, exactly as GoS's engine does.

How the vectors were made

• Skills: the skills_1000 and skills_2000 libraries from DLPenn/graph-of-skills-data at 83a8c59f6d09b849e57a50ed2577e253995d257e, parsed with GoS's parse_skill_document (900-character snippets) at GoS commit 203f60a2c689da055ce1ac351eb3cb9912a3bca7. Each row embeds GoS's node text, SkillNode.from_lists(...).to_str().
• Queries: SkillsBench task instructions from benchflow-ai/skillsbench at 9a1f4dd5f7659f75707435da3ce854b6e48321d1. Each row embeds GoS's query text without LLM rewriting, _fallback_query_schema(instruction).to_query_text(), which is GoS's default (ENABLE_QUERY_REWRITE is off).
• Model: text-embedding-3-large through an Azure OpenAI deployment, computed once in October 2026. Vectors are stored as returned, without renormalization.

Only the public dev queries are included; the task's hidden test queries and their vectors are not published.

Licenses

The vectors are released under MIT. They are derived from the MIT-licensed GoS skill libraries and from SkillsBench instructions (Apache-2.0).