--- title: "Slides: We Cut 94% of AI Coding Tokens With a Local Code Index - Rajkumar Sakthivel, Tesco" category: "slides" video_id: "dRmWYHuIJxM" sourceLabels: ["Public YouTube video frames", "Public YouTube metadata"] --- # Slides: We Cut 94% of AI Coding Tokens With a Local Code Index - Rajkumar Sakthivel, Tesco ## Source Video [We Cut 94% of AI Coding Tokens With a Local Code Index - Rajkumar Sakthivel, Tesco](https://www.youtube.com/watch?v=dRmWYHuIJxM) ## Relationship To World's Fair 2026 These slides are extracted from a public AI Engineer YouTube video connected to World's Fair 2026. Speaker-matched clips are supporting context unless later confirmed as exact session recordings; official livestream recordings are day-level/event-level source material. ## Related Scheduled Sessions - No individual scheduled session mapping has been assigned yet; treat this as an event livestream deck. ## Extracted Slides ![[assets/slides/dRmWYHuIJxM/slide-001.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-001.html) - AI slide classifier: `title_card` confidence `0.99` - Text source: agent_vision. Slide text: > We Cut 94% of Our AI Coding Tokens With a Local Code Index > AI Engineer World's Fair > June 30 – July 2, 2026 • San Francisco > Search & Retrieval Track ![[assets/slides/dRmWYHuIJxM/slide-002.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-002.html) - AI slide classifier: `content_slide` confidence `0.98` - Text source: agent_vision. Slide text: > THE ASSUMPTION > Every AI coding tool we tried had the same assumption: > send as much context as possible. > 45,000 tokens per query > ~5,000 tokens per query ![[assets/slides/dRmWYHuIJxM/slide-003.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-003.html) - AI slide classifier: `content_slide` confidence `0.97` - Text source: advanced OCR `rapidocr-live/bright-screen/opencv-adaptive`. - OCR decision: ready — Dense bullet slide with small explanatory text; OCR is likely more reliable than manual transcription. Slide text: > We optimizedthe modelWeshould have optimized the Context. > Better prompts > "Be concise Only returm relevant code. The model stil received 45k tokens of input. > Model settings > Temperature, top-Pmax lokens control outout shape The 45k input was already sent and biled. > Output compression > Talk like a caveman* Saves:75%.of output (10% of bil) Net Impact - 8%. Wrong 10% > A retrieval layer between codebase and agent > Searchran index, refuri only relevant chunks. 94% fewer tokens. > olara-labs/codecontext-engin? ![[assets/slides/dRmWYHuIJxM/slide-004.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-004.html) - AI slide classifier: `content_slide` confidence `0.98` - Text source: agent_vision. Slide text: > Where your tokens actually go > 90% is input > Output compression > = ~8% off total bill > Input retrieval > = ~61% off total bill ![[assets/slides/dRmWYHuIJxM/slide-005.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-005.html) - AI slide classifier: `content_slide` confidence `0.99` - Text source: advanced OCR `rapidocr-live/bright-screen/contrast`. - OCR decision: ready — Multi-card architecture slide with small labels and metrics across several columns. Slide text: > A local retrieval layer between codebase and agent > Tree-sitter Chunking Retrieval puqAH Compression Chunk Graph Code Confidence Scoring > AST-awaresplits 10 langs Vector+8M25+RRF 94% Signatures+docs 89% CALLS.IMPORTS related Threshold gate filter > Everything runs locally.No cloud,noAPI calls.sqlite-vec +FTS5 +graph in threeSQLite files. > delara-labs/code-context-engine ![[assets/slides/dRmWYHuIJxM/slide-006.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-006.html) - AI slide classifier: `content_slide` confidence `0.98` - Text source: agent_vision. Slide text: > Why not just vector search? > Vector Search > FTS5 (BM25) > RRF Fusion > Recall: 0.78 > Recall: 0.72 > Recall: 0.90 ![[assets/slides/dRmWYHuIJxM/slide-007.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-007.html) - AI slide classifier: `content_slide` confidence `0.98` - Text source: advanced OCR `rapidocr-live/bright-screen/opencv-adaptive`. - OCR decision: ready — Dense two-column slide with small supporting copy and chart labels. Slide text: > The hardest problem wasn'tretrieval It was knowing when retrieval waswrongb > LLM-based scoring Confidence scoring blend > Asked the model to rate relevance. Accurale but +2 3s latency Ycost per query Simitarity Sox > Flxed thresholds Keywords 30x > calike. cosino > O.7 =relevant Broke on short quenies and long quenles. Recency 20x > Simpie heuristic won > no API calls. 50% smilarity +30% keyword K20% recency Adaptive: 0.4ms Lesson: dont reach for an LLM when a welghted avera > elararlabs/code-context-cngine ![[assets/slides/dRmWYHuIJxM/slide-008.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-008.html) - AI slide classifier: `content_slide` confidence `0.99` - Text source: advanced OCR `rapidocr-live/full`. - OCR decision: ready — Benchmark slide with multiple metric panels, small numbers, and a code snippet. Slide text: > BENCHMARK > FastAPI FastAPl:53files,20realquestions,reproducible > Full file baseline Afterretrieval 83,681 tok/q 4,927tok/q 94% > retrieval savings > After compression 523tok/q No cherry-picking.No synthetic queries. > 20questionsadeveloperwouldactuallyask > Recall@10 0.90 $pythonbenchmarks/run_benchmark.py --repo fastapi/fastapi--source-dirfa > elara-labs/code-context-engine ![[assets/slides/dRmWYHuIJxM/slide-009.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-009.html) - AI slide classifier: `content_slide` confidence `0.99` - Text source: advanced OCR `rapidocr-live/bright-screen/contrast`. - OCR decision: ready — Four-panel trade-off slide with small body text in each card. Slide text: > What we're honest about > 94% is against full-file reads Monorepos dilute recall > baseline. savings vs normal behavior are lower.Full-file is our reproducible Claude Code already uses grep and partial reads.Real-world feature-per-file repos hit R=1.00.Focused files retrieve best. On Go's fiber (396 fles),recall dropped to 0.07@10.One- > Embedding model matters What actually worked > bge-small-en-v1.5(384d) is fast,not SOTA.Bigger models lift Simple heunistics over ML.SQLite over specialized DB > cache. over pure vector.Local-first.The boring choices com > elara-labs/code-context-engine ![[assets/slides/dRmWYHuIJxM/slide-010.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-010.html) - AI slide classifier: `content_slide` confidence `0.97` - Text source: advanced OCR `rapidocr-live/right-72/contrast`. - OCR decision: ready — Slide uses a small embedded metrics screenshot with dense numeric text. Slide text: > MEASURABLE > token tracked. Every dollar counted. > my-project.247 queriesLast query 5m ago > 800O088xtokenssaved > Input savings Output savings 48.2ktokens 12.4H tokens $3.62 $186.80 > Total saved 12.4Mtokens $189.62 > Breakdoan: > retrieval chunk coepression outputcoepress* 84% <1% 3% 10.4M$156.60 421.5k 48.2k $3.62 $6.32 > mates.Actuai tokens served vs full-fle basellne,per bucket Dolar coats from five model pricing ![[assets/slides/dRmWYHuIJxM/slide-011.jpg]] - Recreated text/layout view: [open HTML recreation](/assets/slide-recreations/slides/dRmWYHuIJxM/slide-011.html) - AI slide classifier: `content_slide` confidence `0.99` - Text source: agent_vision. Slide text: > The biggest optimization in AI coding isn't the model. It's the context. > 94% fewer input tokens > local no data leaves your machine > MIT free, open source Classification audit: `raw/sources/slide-ai-classification/slides/dRmWYHuIJxM/audit.json` ## Slide-Derived Subjects To Review Subject extraction uses video title, related session titles/descriptions, transcript context, and OCR text when available. OCR is best-effort and should be reviewed against the embedded slide images.