Fix array parameter handling in tool declarations for Gemini and OpenAI #358
Conversation
The Gemini API requires an 'items' field for array-type parameters
in tool declarations to specify the type of array elements. Without
this field, the API returns errors like:
"GenerateContentRequest.tools[0].function_declarations[2].parameters.properties[fields].items: missing field"
This fix adds the required 'items: { type: 'STRING' }' field for
all array parameters when formatting tools for the Gemini provider.
Changes:
- Modified format_parameters method to add items field for array types
- Added comprehensive unit tests for the fix
- Added integration tests to verify proper tool formatting
This resolves issues when using tools with array parameters in
Gemini models, enabling proper function calling with arrays.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <[email protected]>
Removed duplicate test coverage by consolidating tools_array_parameter_handling_spec.rb into tools_spec.rb. The single test file now comprehensively covers: - Array parameters with items field - Mixed parameter types - Required vs optional parameters - Integration test of format_tools method
Similar to the Gemini fix, OpenAI also requires an 'items' field
for array-type parameters in tool/function declarations. Without
this field, tools with array parameters may not work correctly.
This commit adds the required 'items: { type: 'string' }' field
for all array-type parameters when formatting tools for OpenAI.
Changes:
- Modified param_schema method to add items field for array types
- Added comprehensive tests for array parameter handling
- Tests verify both single and multiple array parameters work correctly
finbarr
changed the title
| }.compact | ||
|
|
||
| # Add items field for array types | ||
| property[:items] = { type: 'STRING' } if param.type.to_s.downcase == 'array' |
Previously, array parameters in tool declarations only supported string item types. This was limiting for tools that needed arrays of numbers, booleans, objects, or other types. Changes: - Add item_type parameter to RubyLLM::Parameter class - Update OpenAI provider to use item_type for array parameters - Update Gemini provider to use item_type with proper type mapping - Add comprehensive tests for different array item types This enables tools to declare arrays with specific item types: - Arrays of integers/numbers for numeric data - Arrays of booleans for flag collections - Arrays of objects for complex structured data Addresses feedback from PR crmne#358 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <[email protected]>
|
@crmne Thank you for the feedback! I've updated the implementation to support all array item types, not just strings. Changes made:
The implementation now properly supports arrays of any JSON Schema type, addressing your concern. The gemfile.lock changes are just platform-specific dependencies (ARM64 Darwin) that were automatically added by bundler. The PR has been rebased on the latest main branch and all tests are passing. |
|
Thank you so much @finbarr! Looking at the DSL though I think it doesn't look very RubyLLM-like. I'm thinking it may be worth rethinking the Tool DSL completely and take inspiration from @danielfriis's excellent RubyLLM::Schema. Could be great for 2.0! I'll think about it. |
|
@crmne ok sounds good. My company Shogun is using my fork in production to fix the items type bug for now. As you're thinking about 2.0, one area we've repeatedly run into issues is around observability. It's difficult to get every single raw request/response out of the library right now. |
|
@finbarr have you tried |
|
@crmne yes. But we want every single raw request that gets made, and every single raw response that gets received. The hooks and method returns hide failures in the middle, e.g., tool failure retries. |
|
@crmne it would actually be really nice if there was an easy way to modify requests slightly with an |
|
Could be an interesting hook to add! Feel free to open an issue so we can gauge demand |
Summary
itemsfield for array parameters in tool declarationsGenerateContentRequest.tools[0].function_declarations[2].parameters.properties[fields].items: missing fieldProblem
Both Gemini and OpenAI APIs require an
itemsfield for array-type parameters in tool/function declarations to specify the type of array elements. Without this field, tools with array parameters fail with missing field errors or may not work correctly.Solution
Modified the parameter formatting methods in both providers to automatically add the
itemsfield for all array-type parameters:items: { type: 'STRING' }informat_parametersmethoditems: { type: 'string' }inparam_schemamethodChanges
lib/ruby_llm/providers/gemini/tools.rbto add items field for array typeslib/ruby_llm/providers/openai/tools.rbto add items field for array typesspec/ruby_llm/providers/gemini/tools_spec.rbspec/ruby_llm/providers/openai/tools_spec.rbTest Plan
This resolves issues when using tools with array parameters in both Gemini and OpenAI models, enabling proper function calling with arrays across multiple providers.
🤖 Generated with Claude Code