Web Search

GitAuto provides the model with a web search tool that returns titles, snippets, and URLs for a given query. The model uses this to verify current information about libraries, GitHub Actions versions, API usage, and tool configurations before generating code.

Why This Exists

The model's training data has a cutoff date. It confidently writes code using outdated package versions, deprecated API methods, and non-existent GitHub Action versions. For example, it might use actions/checkout@v3 when v4 is current, or reference a library function that was renamed two versions ago. Web search lets the model verify before writing, reducing hallucinations about the current state of the ecosystem.

Why Models Hallucinate APIs and Libraries

Model training data has a cutoff date. After that date, APIs change, packages get deprecated, and new versions introduce breaking changes. But the model doesn't know its knowledge is stale - it writes code using the API it remembers with full confidence. A model trained before a library's v3 release will confidently generate v2 code that doesn't compile. Training data is a static snapshot from the cutoff date, so the model has no way to know that an API changed after collection. Benchmarks use stable APIs, making this weakness invisible during evaluation.

How It Works

The search tool is defined with a strict JSON schema that requires a search query string. When the model calls it, GitAuto executes the search and returns the top results including titles, snippets, and URLs. The results are snippets only - enough to verify a version number or check if an API exists, but not enough to read full documentation. For that, the model can follow up with the URL fetching tool to read the full page content.

The tool description explicitly tells the model: "NEVER search for repository-specific content - assume the repository is private." This prevents wasted searches for code that won't appear in public results.

Related Features

• URL Fetching - reads the full page content from a search result URL
• Anti-Hallucination Prompts - system instructions that tell the model to verify before writing
• Strict Tool Schemas - ensures the search tool receives well-formed arguments