case_study

Case Study

This directory includes some case analysis. We compare the both method(grep + Claude Context semantic search) and the traditional grep only method.

These cases are selected from the Princeton NLP's SWE-bench_Verified dataset. The results and the logs are generated by the run_evaluation.py script. For more details, please refer to the evaluation README.md file.

• 📁 django_14170: Query optimization in YearLookup breaks filtering by "__iso_year"
• 📁 pydata_xarray_6938: .swap_dims() can modify original object

Each case study includes:

• Original Issue: The GitHub issue description and requirements
• Problem Analysis: Technical breakdown of the bug and expected solution
• Method Comparison: Detailed comparison of both approaches
• Conversation Logs: The interaction records showing how the LLM agent call the ols and generate the final answer.
• Results: Performance metrics and outcome analysis

Key Results

Compared with traditional grep only, the both method(grep + Claude Context semantic search) is more efficient and accurate.

Why Grep Fails

1. Information Overload - Generates hundreds of irrelevant matches
2. No Semantic Understanding - Only literal text matching
3. Context Loss - Can't understand code relationships
4. Inefficient Navigation - Produces many irrelevant results

How Grep + Semantic Search Wins

1. Intelligent Filtering - Automatically ranks by relevance
2. Conceptual Understanding - Grasps code meaning and relationships
3. Efficient Navigation - Direct targeting of relevant sections