README.md

Refinement Chain

This example demonstrates X-Ray analyzing a real LangChain refinement workflow built with runnable composition.

LangChain primitives used:

• ChatPromptTemplate
• ChatOpenAI
• StrOutputParser
• pipe composition (prompt | llm | parser)

Original inspiration:

• ChatPromptTemplate pattern

Model:

• default: gpt-4o-mini
• override with XRAY_LANGCHAIN_MODEL

Requirements:

pip install langchain langchain-openai openai

Run:

python examples/langchain_official/refinement_chain/langchain_official_example.py

This live generation path requires OPENAI_API_KEY.

Replay through the CLI:

python -m cli.main examples/langchain_official/refinement_chain/captured_trace.json

Validate JSON:

python -m json.tool examples/langchain_official/refinement_chain/captured_trace.json

What X-Ray demonstrates:

• how a real LangChain refinement loop evolves across steps
• where early improvement gives way to diminishing structural change
• how the committed trace replays deterministically once saved

Committed artifacts:

• captured_trace.json
• xray_analysis.txt
• ui_example.png

Current observed X-Ray output for the committed fixture:

• peak_step = 1
• waste = 83%
• timeline ends in repeated refinement