pvs-hd-tea / 23ws-LLMcoder

LLMcoder - Practical in winter semester 2023/2024
MIT License
2 stars 2 forks source link

Feedback from 2023-12-05 #49

Closed psaegert closed 11 months ago

psaegert commented 11 months ago
  1. Error Handling Strategies: Instead of relying on a feedback loop, we can also should proactively check for common errors and solve them in the first completion step. We can apply a layered strategy where initial steps involve simpler methods like 1) fine-tuning, 2) offering multiple choice alternatives, and 3) employing tools like mypy checkers. The last resort should be 4) human feedback:

  2. Use of Specific Research and Tools: The feedback mentions reading specific papers 2311.04254 and 2203.14465 related to metastrategies.

  3. Incremental Feedback and Latency Reduction: We should work incrementally to reduce latency in providing feedback, especially for non-LLM (Language Learning Models) information. This could involve tapping LSP information in the background to respond to user changes promptly. For this, we would need a VSCode extension, which is allowed as early as we deem it necessary to advance and solve the problems we disciver

  4. Flexibility in Methods: There's an emphasis on being flexible with the methods used, including employing GPT-4 if it enhances the developer experience. Even time-consuming processes might be acceptable in some cases.

  5. Hallucinations and Contextual Awareness: The tool should be aware of the coding context, like knowing if a function doesn’t exist. It's suggested to use LSP to intercept such instances, which is already used for code highlighting.

  6. Practical Evaluation Over Scientific Studies: The feedback advises using the tool in real-world scenarios to evaluate its effectiveness rather than relying solely on scientific studies. It suggests a diversity of examples across different difficulty levels.

  7. Tool Integration: We are allowed to package the LLMcoder into a VSCode extensions already. However, such integrations shouldn't be the priority unless they add significant value.

  8. Terminology and Presentation: When discussing the tool, terminology should be clear, focusing on terms like 'analyzers' and 'repair'. Also, it's advised to keep the presentation focused, discussing one concept per slide.

  9. Technical Suggestions: There are specific technical suggestions like using tree-sitter instead of AST (Abstract Syntax Tree) and considering the structure of projects through tools like ctags.