CTX: I built a LLM-free code context loader that uses 5.2% of tokens

CTX: LLM-free Code Context Loader

Every AI coding agent faces the same problem: what context to include in the prompt?

Naive approach: dump the whole codebase → 100% of token budget gone.
Smart approach: use BM25 search → misses dependency files, import chains, recent edits.

I built CTX to solve this with a deterministic, zero-LLM approach.

How It Works

CTX classifies every query into one of 4 trigger types and routes it to the right retrieval strategy:

from ctx import CTXRetriever

retriever = CTXRetriever(project_path=".")
result = retriever.retrieve("Where is AuthMiddleware defined?")
# → Trigger: EXPLICIT_SYMBOL → symbol index lookup → returns 2 files

Trigger types:

• EXPLICIT_SYMBOL → exact symbol lookup in structural index
• SEMANTIC_CONCEPT → BM25 hybrid blend
• TEMPORAL_HISTORY → git log + recently modified files
• IMPLICIT_CONTEXT → import graph traversal (finds what the target imports)

Benchmark Results

87 queries, 29 docs:

Metric	CTX	BM25	Dense TF-IDF
R@3	0.862	0.667	0.690
R@5	0.954	0.839	0.805
MRR	0.795	0.611	0.563
Token usage	5.2%	100%	100%

For dependency queries specifically: R@5 = 1.000 (BM25 = 0.400).

COIR Benchmark (CodeSearchNet Python)

Metric	CTX	BM25
R@1	0.720	—
R@5	0.740	—
MRR	0.728	—

Install

pip install ctx-retriever

GitHub: jaytoone/CTX

---

The key insight: query intent classification before retrieval beats generic search for code context loading. The trigger classifier adds ~0ms overhead (pure regex + structural index).

Would love feedback from anyone building RAG pipelines for code.