Idea: Provide File Tree Context to AI for Better Structural Understanding

[mazuralexandru]

Problem:
My C++ project has large utils (>100k tokens). While the large context window allows including them fully, it's inefficient as the AI primarily needs to know about the files' existence and location. This forces a choice between excessive context window usage with largely irrelevant code (including them) or incomplete structural context (excluding them).

Solution:
Provide AI context with the project file tree (paths/names) alongside selected file contents.

Benefit:
This ensures the AI understands the full project structure, enabling it to better handle imports, suggest cross-file function calls, and generate more contextually accurate code, without cluttering the prompt unnecessarily.

Example Scenario:
Consider this project structure:

my_project/
├── src/
│   ├── main.cpp           (5k tokens)
│   ├── core_logic.cpp     (15k tokens)
│   └── network_handler.cpp (10k tokens)
├── include/
│   ├── core_logic.h       (2k tokens)
│   ├── network_handler.h  (1k tokens)
│   └── shared_types.h     (3k tokens)
├── utils/
│   ├── string_helpers.cpp (10k tokens)
│   ├── string_helpers.h   (1k tokens)
│   ├── data_parser.cpp    (120k+ tokens)  <-- Large Util 1
│   └── image_processor.cpp(150k+ tokens)  <-- Large Util 2
├── tests/
│   └── test_core.cpp      (8k tokens)
└── CMakeLists.txt         (1k tokens)

In this scenario, if I'm working on src/core_logic.cpp, I want the AI to know utils/data_parser.cpp and utils/image_processor.cpp exist for accurate include suggestions or understanding potential dependencies, but I don't want their 100k+ tokens consuming the context window unnecessarily. Providing just the tree structure would solve this.

Here's how where I would place the button:

If you like the idea but don't have time, I can try implementing it and send a pull request.

[robertpiosik]

Ideally you should have header file for your large utils, don't you agree? I'm not sure what value plain paths provide to the model.

[mazuralexandru]

Example:
My app generated a large .bin file (200k tokens, not in context). The AI suggested a vague run command (./generatedFile) because it didn't know the exact output path within the project structure. With file tree awareness, it could know the build output location and provide the correct command relative to the project root, like ./main/src/bin/generatedFile.

While manually navigating (cd) to the correct folder works, it's more convenient if the AI provides the complete command runnable from the project root. This enables direct copy-pasting, which is particularly helpful for lengthy or complex commands, eliminating the need to change directories first

[robertpiosik]

Hmm, what if you have multiple binaries? There is no way for the model to know to which refer to. I think the solution is to provide appropriate configuration file in context which describes how the application is built.

I'm researching this and repo prompt includes "file map" of included file. I think it looks cool but models do not care 🤓