Thinking Capture

The sunholo.genai.thinking module extracts and processes "thinking" or "reasoning" content from LLM responses that support extended thinking (e.g., Anthropic Claude with extended thinking, Google Gemini with thinking mode).

Extract from Complete Responses

Simple Extraction

from sunholo.genai.thinking import extract_thinking_simple

text = """<thinking>Let me analyze this step by step.
First, the user wants to know about pricing.
I should check the latest pricing page.</thinking>
Based on the current pricing, the plan costs $49/month."""

thinking, response = extract_thinking_simple(text)
# thinking = "Let me analyze this step by step.\nFirst, the user..."
# response = "Based on the current pricing, the plan costs $49/month."

Detailed Extraction

from sunholo.genai.thinking import extract_thinking, ThinkingContent

text = """<thinking>Analysis...</thinking>
Middle text.
<reflection>Let me reconsider...</reflection>
Final answer."""

contents, cleaned = extract_thinking(text)
# contents = [
#   ThinkingContent(text="Analysis...", tag_type="thinking"),
#   ThinkingContent(text="Let me reconsider...", tag_type="reflection"),
# ]
# cleaned = "Middle text.\nFinal answer."

Supported Tags

Tag	Description
<thinking>...</thinking>	Standard thinking block
<antThinking>...</antThinking>	Anthropic thinking format
<reflection>...</reflection>	Reflection/reconsideration block

Streaming Capture

Separate thinking from response content in real-time during streaming:

from sunholo.genai.thinking import ThinkingCapture

capture = ThinkingCapture()

# Process streaming chunks
for chunk in llm_stream:
    response_text = capture.process_chunk(chunk)
    if response_text:
        # Send to UI (thinking content is filtered out)
        send_to_client(response_text)

# Flush remaining content
remaining = capture.flush()
if remaining:
    send_to_client(remaining)

# Access captured thinking
thinking = capture.get_thinking()
response = capture.get_response()
print(f"Model thought: {thinking}")

Custom Tags

# Capture with a custom tag name
capture = ThinkingCapture(tag="reflection")

State Management

capture.is_in_thinking  # True if currently inside a thinking block
capture.reset()         # Reset for reuse

Anthropic Extended Thinking

Handle Anthropic's native thinking block format (separate content blocks, not tags):

from sunholo.genai.thinking import extract_anthropic_thinking

# response is an Anthropic Message object with content blocks
thinking, response_text = extract_anthropic_thinking(response)
# thinking = content from blocks with type="thinking"
# response_text = content from blocks with type="text"

Streaming Callback Factory

Create a callback that routes thinking and response content to separate handlers:

from sunholo.genai.thinking import create_thinking_callback

thinking_chunks = []
response_chunks = []

callback = create_thinking_callback(
    on_thinking=thinking_chunks.append,
    on_response=response_chunks.append,
)

# Use as a streaming callback
for chunk in llm_stream:
    callback(chunk)