Building a better repository map with tree sitter

robot flowchat

GPT-4 is extremely useful for “self-contained” coding tasks, like generating or modifying a simple function that has no dependencies. Tools like GitHub CoPilot serve these simple coding tasks well.

But making complex changes in a larger, pre-existing codebase is much more difficult, for both humans and AIs. To do this successfully, you need to:

  1. Find the code that needs to be changed.
  2. Understand how that code relates to the rest of the codebase.
  3. Make the correct code change to accomplish the task.

GPT-4 is actually great at making the code changes (3), once you tell it which files need to be changed (1) and show it how they fit into the rest of the codebase (2).

This article is going to focus on step (2), providing “code context”:

To address these issues, aider sends GPT a concise map of your whole git repository that includes the most important classes and functions along with their types and call signatures.

This repository map is now built automatically using tree-sitter to extract symbol definitions from source files. Tree-sitter is used by many IDEs, editors and LSP servers to help humans search and navigate large codebases. Aider now uses it to help GPT better comprehend, navigate and edit code in larger repos.

To code with GPT-4 using the techniques discussed here, just install aider.

The problem: code context

GPT-4 is great at “self contained” coding tasks, like writing or modifying a pure function with no external dependencies. GPT can easily handle requests like “write a Fibonacci function” or “rewrite this loop using list comprehensions”, because they require no context beyond the code being discussed.

Most real code is not pure and self-contained, it is intertwined with and depends on code from many different files in a repo. If you ask GPT to “switch all the print statements in class Foo to use the BarLog logging system”, it needs to see and modify the code in the Foo class, but it also needs to understand how to use the project’s BarLog subsystem.

A simple solution is to send the entire codebase to GPT along with each change request. Now GPT has all the context! But this won’t work for even moderately sized repos, because they won’t fit into the context window.

A better approach is to be selective, and hand pick which files to send. For the example above, you could send the file that contains the Foo class and the file that contains the BarLog logging subsystem. This works pretty well, and is supported by aider – you can manually specify which files to “add to the chat” you are having with GPT.

But sending whole files is a bulky way to send code context, wasting the precious context window. GPT doesn’t need to see the entire implementation of BarLog, it just needs to understand it well enough to use it. You may quickly run out of context window by sending full files of code just to convey context.

Aider also strives to reduce the manual work involved in coding with AI. So in an ideal world, we’d like aider to automatically identify and provide the needed code context.

Using a repo map to provide context

Aider sends a repo map to GPT along with each request from the user to make a code change. The map contains a list of the files in the repo, along with the key symbols which are defined in each file. It shows how each of these symbols are defined in the source code, by including the critical lines of code for each definition.

Here’s a sample of the map of the aider repo, just showing the maps of base_coder.py and commands.py :

aider/coders/base_coder.py:
⋮...
│class Coder:
│    abs_fnames = None
⋮...
│    @classmethod
│    def create(
│        self,
│        main_model,
│        edit_format,
│        io,
│        skip_model_availabily_check=False,
│        **kwargs,
⋮...
│    def abs_root_path(self, path):
⋮...
│    def run(self, with_message=None):
⋮...

aider/commands.py:
⋮...
│class Commands:
│    voice = None
│
⋮...
│    def get_commands(self):
⋮...
│    def get_command_completions(self, cmd_name, partial):
⋮...
│    def run(self, inp):
⋮...

Mapping out the repo like this provides some key benefits:

Optimizing the map

Of course, for large repositories even just the repo map might be too large for GPT’s context window. Aider solves this problem by sending just the most relevant portions of the repo map. It does this by analyzing the full repo map using a graph ranking algorithm, computed on a graph where each source file is a node and edges connect files which have dependencies. Aider optimizes the repo map by selecting the most important parts of the codebase which will fit into the token budget assigned by the user (via the --map-tokens switch, which defaults to 1k tokens).

The sample map shown above doesn’t contain every class, method and function from those files. It only includes the most important identifiers, the ones which are most often referenced by other portions of the code. These are the key pieces of context that GPT needs to know to understand the overall codebase.

Using tree-sitter to make the map

Under the hood, aider uses tree sitter to build the map. It specifically uses the py-tree-sitter-languages python module, which provides simple, pip-installable binary wheels for most popular programming languages.

Tree-sitter parses source code into an Abstract Syntax Tree (AST) based on the syntax of the programming language. Using the AST, we can identify where functions, classes, variables, types and other definitions occur in the source code. We can also identify where else in the code these things are used or referenced.

Aider uses all of these definitions and references to determine which are the most important identifiers in the repository, and to produce the repo map that shows just those key lines from the codebase.

What about ctags?

The tree-sitter repository map replaces the ctags based map that aider originally used. Switching from ctags to tree-sitter provides a bunch of benefits:

Future work

You’ll recall that we identified the 3 key steps required to use GPT to complete a coding task within a large, pre-existing codebase:

  1. Find the code that needs to be changed.
  2. Understand how that code relates to the rest of the codebase.
  3. Make the correct code change to accomplish the task.

We’re now using tree-sitter to help solve the code context problem (2), but it’s also an important foundation for future work on automatically finding all the code which will need to be changed (1).

Right now, aider relies on the user to specify which source files will need to be modified to complete their request. Users manually “add files to the chat” using aider’s /add command, which makes those files available for GPT to modify.

This works well, but a key piece of future work is to harness the power of GPT and tree-sitter to automatically identify which parts of the code will need changes.

Try it out

To code with GPT-4 using the techniques discussed here, just install aider.

Credits

Aider uses modified versions of the tags.scm files from these open source tree-sitter language implementations: