Denspi

The interesting bet in Denspi is architectural. Most open-domain QA systems of its era split the job: a retriever narrows millions of documents to a handful, then a reader extracts the span answer. That pipeline caps your latency and, more subtly, caps your recall — if the retriever misses the passage, no reader can recover it. Denspi removes the retriever entirely by precomputing an index over every phrase in Wikipedia, so at query time you do one nearest-neighbour lookup against a joint dense-sparse representation. The dense side is a BERT-based phrase embedding; the sparse side is TF-IDF vectors from DrQA. Faiss handles the vector search, and the paper's headline number is reading the whole of Wikipedia in about 0.5s on CPUs — roughly 58x faster than retrieve-and-read approaches, which is what makes long-tail answers practical rather than theoretical. The practical picture is less glamorous. You need at least 4 CPU cores, 30 GB of RAM and 1.5 TB of SSD for the dump, and the download itself is the first real barrier. Setup assumes a Conda environment on Python 3.6 because Faiss can't be installed via pip, and DrQA has to be installed manually first — the README even notes you'll want java-jdk for it. Two separate requirements.txt files apply, one for the PyTorch question-encoding server and one in the open/ folder for the search server and demo. The payoff is that once files are local, serving takes under a minute. If you intend to train rather than use the pretrained model, the hardware bar jumps: 4x P40 (24 GB) GPUs for the stated batch size, about five hours per epoch over three epochs on SQuAD v1.1 — roughly 16 hours — followed by a finetuning pass with negative samples drawn from other documents. The codebase is a genuine research repository: 250 commits, 200 stars, 23 forks, nine branches, six open issues and five open pull requests. Custom index building code is present in the open/ folder with an updated readme, so indexing a corpus other than Wikipedia is a supported path, not a hack. The honest limitations are the ones the seed notes and the README confirms. Answers are tied to a fixed Wikipedia snapshot. Accuracy degrades on out-of-domain questions. There's no conversational or multi-turn layer. Nothing here is productized: no SLA, no roadmap, no support contract, and the environment pins are from 2019. Compare it to ColBERT or DrQA if you want the same retrieval ideas with more surrounding activity, or to a managed QA API if you want something you don't have to operate. Denspi's value is that it makes a strong, reproducible argument that phrase retrieval is viable at web scale — for a researcher, that's worth the 1.5 TB.