feat: Add Gemini "Thinking" support to GoogleGenAIChatGenerator

[vblagoje]

• fixes Google GenAI - reasoning support #2176

Why:

Introduces support for Gemini "thinking/reasoning" .Models can now show their reasoning process, maintain thought context across multi-turn conversations, and use configurable thinking budgets.

What:

• Added thinking_budget configuration via generation_kwargs with model-specific validation
• Implemented parsing of thought parts into ReasoningContent objects
• Extended streaming aggregation to handle reasoning deltas while reusing generic aggregator
• Implemented fail-fast error handling for unsupported models
• Added integration tests for thinking feature and error scenarios

How can it be used:

from haystack_integrations.components.generators.google_genai import GoogleGenAIChatGenerator
from haystack.dataclasses import ChatMessage

# Enable thinking with dynamic or explicit budget
generator = GoogleGenAIChatGenerator(
    model="gemini-2.5-pro",
    generation_kwargs={"thinking_budget": -1}  # -1 for dynamic, or specific token count
)

messages = [ChatMessage.from_user("Explain why the sky is blue")]
response = generator.run(messages=messages)

# Access reasoning content
message = response["replies"][0]
if message.reasonings:
    for reasoning in message.reasonings:
        print(f"Thought: {reasoning.reasoning_text}")

# Check thinking tokens used
print(f"Thinking tokens: {message.meta['usage'].get('thoughts_token_count', 0)}")

How did you test it:

• Added integration tests for thinking feature with gemini-2.5-pro
• Added fail-fast tests for unsupported models

Notes for the reviewer:

• _aggregate_streaming_chunks_with_reasoning extends the generic aggregator (40 lines vs 78 if duplicated)
• Model detection may need updates as new models are released

[anakin87]

My impression is that handling thinking + tool calls have some peculiarities to take into consideration
https://ai.google.dev/gemini-api/docs/function-calling?example=meeting#thinking
https://ai.google.dev/gemini-api/docs/thinking#signatures

[vblagoje]

My impression is that handling thinking + tool calls have some peculiarities to take into consideration https://ai.google.dev/gemini-api/docs/function-calling?example=meeting#thinking https://ai.google.dev/gemini-api/docs/thinking#signatures

Ok let me check

[anakin87]

Please add unit tests to cover the new methods and code paths introduced in this PR

[anakin87]

From the CI, I see this coverage

Name                                                                                  Stmts   Miss Branch BrPart  Cover   Missing
---------------------------------------------------------------------------------------------------------------------------------
src/haystack_integrations/components/generators/google_genai/chat/chat_generator.py     338     53    156     32    80%   77-78, 99-106, 108->118, 110->118, 123->128, 136-137, 141-142, 145-150, 161-163, 184-188, 195-205, 228-229, 275->309, 278->309, 305->279, 542, 564->568, 611-616, 631, 670->676, 672->671, 677-681, 684->666, 692, 697, 737-739, 772-774, 850-851, 862, 866, 904-905, 959-960, 971, 975, 1013-1014

Could you please add some additional unit tests to cover what changed in this PR?
For example, _convert_message_to_google_genai_format...

[vblagoje]

Nice nice, on it! One sec 🙏

[vblagoje]

@julian-risch @anakin87 is this one good to go? LMK

[anakin87]

Ok for me.
My impression is that @julian-risch wants to review as well.

[vblagoje]

@julian-risch would you please take a look - I'd like to integrate this one and release a new version soon

[vblagoje]

@julian-risch I think @anakin87 is a very good reviewer so I took the liberty to move this PR forward.

[vblagoje]

Included in https://pypi.org/project/google-genai-haystack/2.1.2/ release