We Cut 94% of AI Coding Tokens With a Local Code Index
summary
Claims a 94% reduction in tokens by inserting a local retrieval layer between codebase and agents and sending only relevant snippets instead of full files. Uses AST-aware chunks (tree-sitter), combined vector search plus exact keyword matching, and a lightweight relationship layer to follow execution across files. Shows numbers: 45,000 tokens typical query with ~5,000 useful, FastAPI test drops ~83k → 4.9k (523 with compression) with ~90% recall. Runs locally, uses a 50% meaning 30% keyword 20% recency heuristic scoring that executes in ~0.4 ms.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 1:06AI coding token waste
- 2:32Output compression limits
- 3:15Input tokens save costs
- 3:38Local code search layer
- 4:43Two searches beat one
- 5:48Judging search results
- 6:34FastAPI test case savings
- 7:39Limitations and trade-offs
- 8:21Simple choices
- 8:44Shared index for AI tools
- 9:29Measured savings
- 10:11Model choice vs input costs
related talks
31mAI Coding Agents Don't Need Bigger Models. They Need Better Tools.
45mAgentic AI Coding Practices for Speed with Quality
22mSponsored Demo: Cut the Bill, Not the Brains. AI Coding Agents That Ship More and...
30mBeyond the Hype: Using AI Coding Agents Today
30mHow AI Agents Will Change the Future of Programming
28mAI Migration at SCALE: Make Complex Code Tasks 10X Faster (The Blueprint)
26mCoding with AI
43mAI accelerated legacy modernisation