Search blog posts, product pages, and docs by title, topic, or keyword. Try “RAG”, “compliance”, or “CI/CD”.
Posts in categories: "Research" (14 posts)
We put thirteen models and a keyword rule on the two safety decisions a medical-practice chat makes before it answers. One model caught every message from a writer at risk, and flagged every safe message too. What the board can rank, what it cannot, and why half of it is sealed.
Read More about Every Crisis Caught, Every Message FlaggedThree vector indexes tied at about 0.90 nDCG@10 on a public benchmark. The chat path built on them did not, and the whole gap was questions that reached the model with no context.
Read More about The Index Was Fine: Measuring a RAG Path End to EndWe swapped a 70B chat classifier for a typed-decision model and got 7x the speed at higher accuracy. The result worth writing down was the one where both models scored exactly the majority-class baseline.
Read More about The Question a Classifier Cannot AnswerSeven of eight major agent benchmarks were driven to near-perfect scores without solving a task — and every one of those runs would have passed a signature check. The unit of trust is not the number. It is the manifest, and a manifest has to be able to state its own weakness.
Read More about What a Benchmark Has to Prove About ItselfA Rust daemon on one MacBook published a dozen sites a day. The same pipeline now runs on Cloudflare, unattended, and publishes about a thousand.
Read More about WWW-RAG: Making the Open Web Chattable, From One MacBook to CloudflareAn e-commerce assistant linked products correctly but recommended the wrong ones. The fix was discovering that decoration, ranking, and recommendation are three different problems — and that prominence is only ever a tiebreaker on relevance.
Read More about Prominence Is Not Relevance: Teaching a Storefront Assistant What to RecommendWe swapped OpenParse for LiteParse in our RAG ingestion pipeline. The headline is 20–50× faster parsing. The useful part is the four ways we got it wrong first.
Read More about We Made Our RAG Pipeline Parse PDFs 20–50× FasterOpen-source Headroom compresses RAG context with an ONNX model. We wired it into our pipeline and raced it against a 50-line in-process extractor. Neither won outright.
Read More about We Tested Headroom Against Our EXIT RAG Compressorz-lab's DFlash drafter on our QLoRA fine-tune captured 92% of the published speedup with no retraining. ~15x faster, ~4x cheaper in prod.
Read More about Speculative Decoding for Free: Pairing DFlash with our DFO-Tuned Gemma 4 31BTwo natively-trained 1-bit LLMs converge on the same activation anomaly without ever sharing weights. A note on convergence under pressure.
Read More about The Two Models That Never Met. Both Measured at the Same Depth.Two natively-trained 1-bit LLMs lose the four-stage circuit that organizes fp16 transformers. Behavior survives; structure dissolves.
Read More about When the Circuit DissolvesA 200-item RAG arena tied at the mean, but two LLM judges only agreed at Spearman ρ=0.55. They aren't measuring the same thing.
Read More about Inside the RAG Arena: When the Judges Don't AgreeA vIndex patch deletes 'Paris is the capital of France' from a frozen LLM. The receipt is a 100-byte JSON file with a SHA-256 checksum.
Read More about Deleting Paris from a Language ModelEvery modern LLM converges on the same four-stage circuit. We map the architecture and what its absence in 1-bit models means.
Read More about The Architecture Every Language Model Converges To
