v2.8.0

This release promotes the Ollama web search integration to production, with significant improvements to prompt quality, dependency management, and validation flexibility.

New Features:

• Ollama Web Search Tool: Added support for Ollama as a web search provider, expanding the available search options beyond DuckDuckGo and builtin provider tools.
  • Introduced OllamaWebSearchDeps and ollama_web_search_multi function for web search.
  • Added OllamaSearchSettings configuration class with API key validation.
  • Updated web_search_tool option to include 'ollama' alongside 'builtin' and 'duckduckgo'.
  • Refactored agent initialization to use AsyncExitStack for proper resource cleanup.

Refactoring & Improvements:

• AgentDeps Architecture: Introduced a new AgentDeps data model to centralize dependencies passed to Pydantic AI agents. This replaces the loose Any type with a typed, extensible container holding video_duration_ms and optional ollama_search fields.
  • Updated RateLimitedAgentWrapper to use typed AgentDeps instead of Any.
  • Refactored ollama_web_search functions to extract dependencies from ctx.deps.
  • Updated main.py to create and configure AgentDeps with Ollama search dependencies.
• Rate Limiting Migration: Implemented hook-based rate limiting.
  • Moved rate limiting logic from the run() method to model request hooks via a _rate_limit hook function.
  • Added token calculation helper _calculate_tokens() for accurate rate limit tracking.
  • Improved error message formatting to show human-readable durations.
• Agent Constructor Updates: Refactored both Gemini-cli and default Agent initialization to include capabilities (hooks) and deps_type (AgentDeps) to ensure that rate limits are enforced across all execution paths.
• Deferred Response Validation: Refactored response handling to store AI responses in a local variable, validate against video duration, and only assign to job.response if validation passes. This prevents storing potentially invalid responses on the job object.
• In-Memory Caching for Ollama Search: Added a cache to OllamaWebSearchDeps that stores search results keyed by a normalized query string (punctuation removed, case-folded). This avoids redundant API calls for equivalent queries like "Hello, world!" and "hello world", reducing latency and API costs.

Prompt Engineering:

• Lyrics Detection (v7): Updated LYRICS_PROMPT_VERSION to 7 with refined search rules that prioritize original composer names and native titles. Added concrete search query examples for Japanese covers, English covers, and original songs to improve search accuracy.
• Ollama Search Prompt (v6): Updated LYRICS_PROMPT_VERSION to 6 with refined instructions to prevent excessive search queries. Added explicit rules against query spamming, restrictive quotation marks, and guidance to use native titles for better search results.

Fixes & Improvements:

• Configuration Validation: Updated config validation to only enforce an Ollama API key when the web_search_tool is set to "ollama" AND the thread lyrics flag is enabled (lyrics > 0). This allows for flexible configurations where Ollama search is configured but not actively used.
• Gemini CLI Web Search Warning: Previously, the Gemini CLI branch silently discarded builtin_tools and function_tools when use_web_search was True, meaning the configured web_search_tool was silently ignored. Now a logfire.warn() is emitted explaining that the tool will be skipped, and the agent is created without web search.
• Developer Experience: Added the Ollama API key registration link (https://ollama.com/settings/keys) to validation error messages, making it easier for users to configure their API keys.
• Documentation: Improved docstring clarity and type annotations across config, data models, models, job runner, main, ollama web search, prompt, and video modules. Updated to use modern Google-style format without redundant type hints in Args/Returns sections.
• Documentation: Fixed README to reference the correct split setting (max-seconds vs max_minutes).
• Type Safety: Added proper type annotations to validator methods and fixed field descriptions for rate limiters.

What's Changed

• revert: remove LyricsGenius integration, restore DuckDuckGo search by @FlippFuzz in #6
• Ollama search by @FlippFuzz in #7

Full Changelog: v2.7.0...v2.8.0