AgentSkillsCN

procgen-search

利用语义相似度搜索已索引的ProcGen语料库。对于给定的查询，返回最相关的文档路径。

SKILL.md
--- frontmatter
name: procgen-search
description: Search the indexed procgen corpus using semantic similarity. Returns the most relevant document paths for a given query.
allowed-tools:
  - Bash
  - Read

procgen-search (Vector Embedding Search)

Use this skill to find relevant procgen papers/documents for a given query.

Why vector search?

The vector search approach is cleaner than a "librarian agent" because it's deterministic and fast. Rather than having an LLM guess which document is relevant, we compute semantic similarity directly using sentence-transformers embeddings.

Prerequisites

Before first use, ensure the corpus is indexed:

bash
python .claude/skills/procgen-search/scripts/index_corpus.py

Required packages: sentence-transformers, numpy

Usage

This skill reads $ARGUMENTS as the search query.

Example: /procgen-search WFC terrain generation with constraints

Procedure

  1. •

    Run the search script with the query:

    bash
    python .claude/skills/procgen-search/scripts/search_corpus.py "$ARGUMENTS" --json --top_k 3
  2. •

    Parse the JSON output to get the most relevant document path(s).

  3. •

    Return the results to the caller. The output includes:

    • •path: Full path to the document
    • •filename: Document filename
    • •similarity: Cosine similarity score (0-1)
    • •preview: First 500 chars of the document

Integration with /rlm

After finding relevant documents, chain to the /rlm skill:

code
/rlm context=<best_path> query=<original_user_query>

Reindexing

When new papers are added to ./corpus/, reindex:

bash
python .claude/skills/procgen-search/scripts/index_corpus.py