This is a real scan output from our reference corpus. Findings, dollar savings, and confidence bands render exactly as they would on your own repo. No signup needed to read this page — signup gives you the downloadable agent-instructions.md your coding agent applies.
summarization_chain.pySummarization chain uses gpt-4 for one-sentence summaries. gpt-4o-mini is ~10x cheaper with equivalent quality on this task.
FixChange model="gpt-4" to model="gpt-4o-mini" in ChatOpenAI constructor. Reduce max_tokens from 2000 to 150.
retrieval_qa.pyRetrieval QA chain uses k=20 with stuff chain type, sending 20 full documents per query. No cache on repeated queries.
FixReduce k from 20 to 5. Add an LRU cache on answer() keyed by question hash. Consider map_reduce chain type for long contexts.
agent_tools.pyAgent uses max_iterations=20 on gpt-4 with ZERO_SHOT_REACT_DESCRIPTION. A failing query can cost 20x the expected tokens before converging.
FixLower max_iterations to 6. Switch to OpenAI Functions agent type for reliable tool calls with fewer iterations. Enable Anthropic prompt caching on tool schemas if switching to Claude.
translation_service.pyTranslation service iterates texts one at a time instead of batching through a single chain.batch() call.
FixReplace the list comprehension in translate_many with chain.batch(). Switch model to gpt-4o-mini for short-form translation.