LLM Extreme Context

Configuration

This project uses a settings.json file for configuration. If the file does not exist it will be created automatically with default values. The settings.example.json file is kept in sync with the defaults whenever the tools run, so copying that file is an easy way to start customizing your own settings. The extractors handle Python, HTML, Markdown, JavaScript, and TypeScript source files using tree-sitter for the latter. CSS variants, modern JavaScript modules (.mjs, .cjs), .tsx, Vue, Svelte, and SVG files are also supported using text-based extraction. JSON, YAML, and plain text files are parsed so every piece of content can be searched.

Installation

Install dependencies with:

pip install -r requirements.txt

Workflow

1. extract – parse source files and build a call graph
2. embed – generate embeddings for each node
3. query – interactively search the embeddings
4. inspect – optional graph statistics and visualisation

If sub_question_count is greater than 0, an LLM call breaks your question into that many sub-queries before searching the embeddings. This helps find relevant code from multiple angles.

During querying, the LLM may request additional function summaries before providing a final answer. Each response is a JSON object with either {"response_type": "functions"} requesting more code by name or {"response_type": "info"} containing the final summary.

Run the tests with pytest to ensure everything works:

pytest -q

Settings Options

• Settings are grouped into categories for easier navigation:

• LLM_model – api_key, api_type, local_path

• api_settings – temperature, top_p, max_output_tokens

• encoder_model_path – location of the sentence transformer model. If left blank, the small sentence-transformers/all-MiniLM-L6-v2 model will be downloaded automatically.

• paths – output_dir

• query – top_k_results, use_spellcheck, sub_question_count

• context – context_hops, max_neighbors, bidirectional

• extraction – allowed_extensions, exclude_dirs, comment_lookback_lines, token_estimate_ratio, and minified_js_detection options

• visualization – parameters controlling call graph rendering; set auto_visualize to true to automatically save a PNG after extraction

• embedding – encoder_model_path

The extraction step relies on crawl_directory which automatically skips files listed in .gitignore and only processes paths with extensions from allowed_extensions.

When bidirectional is True, context expansion walks both callers and callees. Set it to False to only follow outgoing calls.

Files

• main.py: Unified CLI and interactive workflow
• graph.py: Source extraction and call graph helpers
• embedding.py: Generates embeddings and FAISS index
• query.py: Runs searches and prepares prompts
• prompt_builder.py: Formats results for the LLM
• llm.py: Gemini API helpers
• interactive_cli.py: PromptToolkit-based input dialogs
• config.py: Loads and saves settings
• settings.json: Your local configuration (not tracked by git)
• settings.example.json: Template configuration file

Usage

The settings.json file is automatically loaded by all scripts. If the file doesn't exist, it will be generated with defaults the first time you run the tools. Invoke the CLI with python main.py <command>:

python main.py extract /path/to/project
python main.py embed my_project
python main.py query my_project --problem "bug"
python main.py inspect my_project

Running main.py without arguments starts an interactive loop. It will prompt for the project path and problem statement. The LLM then guides an iterative search by requesting additional functions until it can provide a final summary. Use neighbors <n> to inspect graph neighbors of a result. Prompt history is saved under ~/.full_context_history/.

Spellcheck and Sub-Queries

Two optional features help refine search queries:

• use_spellcheck – when true, the query is corrected using SymSpell before searching. A small dictionary is built from item names automatically, but you can also download a larger frequency dictionary from the SymSpell repository and place it in the project root.
• sub_question_count – when greater than 0, your question is first sent to the LLM to produce that many sub-queries. Their embeddings plus the original query are averaged together before searching. The default value is 0 (off).

Enable these options in settings.json under the query section.

API Settings

The api_settings section controls parameters sent to the LLM API. Adjust temperature, top_p, and max_output_tokens to tune response style and length. The default max_output_tokens is 5000.

Debug Logging

All tools output progress information using the logging module and tqdm progress bars. Run scripts with the PYTHONLOGLEVEL=DEBUG environment variable to see verbose details while files are scanned and embeddings are generated.