LLM Wiki — Semantic Linker Skill (wiki_semantic_link)

Resolving script paths (read first): Commands below invoke scripts as <BIN>/X.py (and a few as <SKILLS>/...). Resolve these to absolute paths once before running anything:

>

- <SKILL_DIR> = the directory this SKILL.md lives in. - <SKILLS> = the skills/ folder containing this skill = <SKILL_DIR>/.. - <BIN> = the bin/ folder beside it = <SKILL_DIR>/../../bin

>

Do not hardcode a fixed prefix like .agents/bin or ../bin: shell relative paths resolve against the current working directory (usually the topic root), not this skill's location. Once resolved, <BIN> is typically .agents/bin when invoked from the hub root, or .claude/bin from inside a topic directory.

This skill scans all markdown files within the wiki/concepts/ directory, extracts their text, and generates embeddings using a local Ollama model (configured in config.yaml, default: qwen3-embedding:0.6b). It then calculates pairwise cosine similarity between all concepts and automatically injects bi-directional Obsidian-style links ([[Concept Name]]) for pairs that exceed a given similarity threshold.

Execution

This skill is executed via the Python script located within the skill directory.

Crucial Parameter Selection: Depending on the objective, you MUST use the correct parameters:

1. Normal Link Injection: If the user just wants to inject semantic links (default behavior):

python <SKILLS>/wiki_semantic_link/semantic_linker.py <TOPIC_DIR> Model, threshold, and Ollama URL are read from config.yaml. Override via CLI flags if needed (e.g. --threshold 0.80 --model <model>).

1. Deduplication Scan (Auto-Merge): If the user is running the wiki_concept_sync pipeline to merge duplicates:

python <SKILLS>/wiki_semantic_link/semantic_linker.py <TOPIC_DIR> --dedup-only --auto-merge

Prerequisites: Ensure the embedding model is available locally via ollama pull <model> (check config.yaml for the configured model name). Requires numpy and scikit-learn Python packages.

Internal Phases

Phase 1: Preparation & Backup

1. Identifies the wiki/concepts/ directory within the given <TOPIC_DIR>.
2. MANDATORY: Copies all .md files to wiki/concepts/.backup/ before proceeding, ensuring no data loss in case of a bad threshold.

Phase 2: Embedding Generation

1. Reads all concept markdown files.
2. Strips existing YAML frontmatter and internal links to focus on the pure definition/content for unbiased embeddings.
3. Calls the local Ollama instance (configured in config.yaml) to generate vector representations.

Phase 3: Similarity Calculation & Filtering

1. Uses scikit-learn to compute a Cosine Similarity matrix for all generated vectors.
2. Filters out identical pairs (similarity = 1.0, i.e., self-links) and pairs below the defined threshold (e.g., < 0.85).

Phase 4: Idempotent Injection

1. For every highly-related pair (A, B) (where score >= threshold), checks if A.md already contains [[B]].
2. If not, it safely injects it at the bottom under the heading ## 语义关联 (Semantic Links).

Note: This skill no longer handles merge suggestions. Deduplication is strictly handled by the `wiki_concept_sync` skill.