Code Completion Quality - Ideas

[wsxiaoys]

Under construction

Since the release of version 0.13, the core functionality of code completion has become largely stable. This marks an opportune moment to consolidate current implementation and begin exploring for more enhancements

Prompt construction

Tabby generates code completion requests using the following information:

• prefix: the code that precedes the cursor.
• suffix: the code that follows the cursor.
• lsp definitions: for the K identifiers located before the cursor, their definitions are retrieved using the IDE's internal LSP.
• recently changed content: lines that have been recently modified (potentially from different files) within the current workspace.
• repository-level relevant code snippets: relevant snippets are identified using the prefix and then scored with a combination of BM25 and embedding search algorithms.

It is important to note that prefix and suffix are guaranteed to be included in the LLM inference request, provided they exist. However, lsp definitions, recently changed content, and repository-level relevant code snippets are subject to the limitations of the context window, which is currently set to 1536 tokens. The priority in which these elements are filled corresponds to the order listed above.

To ensure low latency, we are currently using a relatively conservative max decoding tokens, 64.

Ideas

Server

• make context length configurable, allowing larger context being sent to LLM feat(config): support configure input length #2649
• make max decoding tokens configurable, or set it dynamically in a smart way chore: make max_output_tokens configurable #2683
• Completion results are worse with Windows line endings (CRLF) #3279

Agent

• add recent opened files to code completion requests #2664
  • About the IDE plugin's ability to read nearby tab context #2108 (comment)

IDE / Extensions

• IntelliJ Plugin: support collecting declarations #3421
• VIM / NeoVIM: support LSP definitions

[5bug]

Agent：
Can the number of lines of context code support a dynamic mechanism that can take the lines of a complete definition or function? Would this enhance the model's understanding ability?

[5bug]

IDE / Extensions:
Clipboard content is supported as a recommended reference

[lampewebdev]

Completion preview:
configure a min length for code shown to complete.
Sometimes when i press enter and want to press "TAB". Tabby in the same moment suggest }), which then i have to remove. So it would be nice to not suggest completions that are less then 3 chars or something I can configure.