- Always use actual code as query input, not natural language descriptions
- Prerequisite:
symbols.dbmust exist at project root (built viaarmyknife searchx index)
Workflow
- User provides code as argument (e.g.,
/similar-code <pasted code>) - Run the search command:
armyknife searchx query '<code>' --limit 10 --database symbols.db
- Present results as a table (symbol, kind, file, line, similarity)
- Read top matches to show code context
Interpreting Results
| Similarity | Meaning | |------------|---------| | 0.8+ | Very similar (likely same pattern or implementation) | | 0.6-0.8 | Related (similar structure or purpose) | | <0.6 | Weak match |
Limitations
| Input Type | Result Quality | |------------|----------------| | Actual code | Good (designed for code-to-code similarity) | | Natural language | Poor (embedding model is code-focused) |
Code is embedded using embeddinggemma:300m (768-dim) via Ollama.
微信扫一扫