llm-diet 0.1.2

Deterministic codebase context retrieval for LLMs

context-engine

Cut your AI coding costs by 99%. Inject only the code that matters.

Record your own: bash demo/record_demo.sh

Benchmark (this repo, 185 nodes)

Query	Baseline Tokens	With CE Tokens	Reduction %	Nodes	Files Selected	Time (ms)
fix authentication bug	46,661	434	99.1%	3	intent.py, validator.py, patcher.py	187
add a new API endpoint	46,661	106	99.8%	1	ranker.py	203
how does the database connection work	46,661	278	99.4%	2	intent.py, retrieval.py	219
debug memory leak	46,661	428	99.1%	3	intent.py, cli.py, compressor.py	172
add input validation	46,661	0	100.0%	0	no match in this repo	125
explain the caching logic	46,661	418	99.1%	3	intent.py, pruner.py, retrieval.py	172
fix error handling	46,661	479	99.0%	3	intent.py, validator.py, patcher.py	187
add logging to the pipeline	46,661	58	99.9%	1	cli.py	141

46,661 tokens → 275 tokens average. Same accuracy. 176ms overhead (small repo).

Benchmark (external repo — fastapi, 946k token codebase)

Query	Baseline Tokens	With CE Tokens	Reduction %	Nodes	Files Selected	Time (ms)
fix authentication bug	946,210	87	>99.9%	1	api_key.py	359
add a new API endpoint	946,210	120	>99.9%	1	api_key.py	359
how does the database connection work	946,210	130	>99.9%	2	tutorial001_an_py310.py, param_functions.py	391
debug memory leak	946,210	436	>99.9%	5	applications.py, test_arbitrary_types.py, …	1062
add input validation	946,210	244	>99.9%	3	utils.py, tutorial004_py310.py, …	375
explain the caching logic	946,210	114	>99.9%	1	test_security_scopes_sub_dependency.py	313
fix error handling	946,210	221	>99.9%	3	tutorial003_py310.py, tutorial002_py310.py, …	406
add logging to the pipeline	946,210	136	>99.9%	1	tutorial002_an_py310.py	328

946,210 tokens → 186 tokens average. Repo never seen before. 449ms overhead.

How it works

User prompt → context-engine → top 5 relevant functions → Claude sees 275 tokens
                                        ↑
               (instead of your entire codebase at 46,661 tokens)

AST-parses your repo into a call graph. When you send a prompt, it scores every function by keyword relevance + call graph centrality and injects only the top matches — before Claude starts reasoning.

No embeddings. No vector DB. No LLM calls in the retrieval path.

Install

pip install llm-diet
context-engine index .

Use with Claude Code (automatic)

claude mcp add context-engine -- python -m context_engine.mcp_server

After this, every prompt you send to Claude Code is automatically prefixed with the minimal relevant context from your codebase.

Use as CLI

context-engine query "fix the auth bug"
context-engine apply "add input validation to the login endpoint"

Why not just use RAG / embeddings?

No LLM calls, no vector DB, no setup.

context-engine uses AST parsing + call graph traversal. It's deterministic — same query, same graph, same result every time. Works offline. Runs in ~176ms (small repo) / ~432ms (large repo like FastAPI). Embeddings need a model call just to retrieve context. We don't.

Feature	context-engine	code-review-graph
Languages	Python, JS, TS, JSX, TSX	23 languages
Dependencies	tree-sitter only	tree-sitter + SQLite + more
Context injection	UserPromptSubmit hook (Claude Code) / IDE rules (Cursor, Windsurf)	MCP server
Autonomous apply	✅ plan → diff → validate → patch	❌
Setup	pip install + index	pip install + build
Incremental updates	auto on file save via context-engine watch	auto on file save

We do less. What we do, we do surgically.

* Full token reduction applies to Claude Code. Cursor/Windsurf receive static rules files that guide the AI to prefer provided context — dynamic injection coming in a future release.