README

This is an early POC to probe the possibility to use tree-sitter to automate vendor migrations. As an example codebase we use exhaustive scenarios of the OpenAI TypeScript SDK and aim to automate the migration of all GPT-3.5 calls to Mistral's SDK.

The goal is that for >80% of cases we can automatically integrate Mistral correctly and for the remaining 20% we can provide a clear and correct todo list for the developer to complete the migration.

The project is in a very, very early stage to find obvious roadblocks early on. See the Development section for more details on this.

Usage

The repo exposes a CLI to run the parser on a given directoy. To use it on the example codebase, run:

cargo build && target/debug/cmod -p ./examples/openai-starter/src

For a more systematic probing of the functionality, you can run the tests:

cargo test

Note: The example codebase(s) to vendor-integrate are in the example directory.

Development

The to-do list to achieve the above stated objective (migrate GPT3.5 to Mistral) is as follows:

find all target files, i.e. importing 'openai' (see: get_target_files)
search for all createCompletion calls (probably doable)
rewrite to use mistral's version of that (more research needed)
- Investigate how to use tree-sitter to rewrite the AST; check ast-grep
insert the import statement for mistral sdk file (trivial)
create mistral sdk file (trivial)
add mistral as a dependency and prompt to run npm i. (trivial)

Notes on Step 1

Finding the target files is rather trivial (excluding extrem edge cases like dynamic imports) and are essentially compiler correct. Core implementation in the parser.rs file via the get_target_files function.

Notes on Step 2

Once you have the target files, targeting the correct function calls for change like createCompletion has one additional complexity: you need to resolve by entity from the imported openai module. This should be achievable via tree-sitters tagging system. This is not implemented yet though. Alternative approaches would be LSP/LSIF or sourcegraph's improvement over LSIF aka. SCIP, which, unlike tree-sitter, were designed for codebase-level (instead of file-level) traversal. However, tree-sitter is significantly faster, efficient, and easier extensible than SCIP/LSIF.

Notes on Step 3

This step is the most uncertain. Not so much if it is doable (liberaries like ast-grep implemented this in a general manner with a <1k LOC) but more so if you try to bend tree-setter too much into doing something that it is not designed for. A lot of inspiration can be taken from other codemod/rewrite tools. For our use-case a declarative API might be sufficient and preferred.

Notes on Step 4-6

This is trivial stuff and essentially basic codegen (rather than codemod) stuff. For a fully working POC this would be necessary to implement though.