Denspi

Denspi's contribution is conceptual as much as practical: it removes the reader entirely. Rather than retrieving documents and then extracting an answer span, DenSPI enumerates, embeds and indexes every phrase in Wikipedia (60 billion) so that open-domain QA becomes a pure phrase retrieval problem. The index mixes dense phrase embeddings from a BERT-based encoder with paragraph-level sparse TF-IDF vectors from DrQA, all served through Faiss. The payoff the authors report is latency — reading the entire Wikipedia in roughly 0.5 seconds on CPUs, at least 58x faster inference than retrieve-and-read models — which is what makes long-tail answers reachable in real time. What you actually get in the repository: a PyTorch-based server that maps an input question to a BERT embedding and returns it as JSON over a REST port; an hdf5-based phrase index with a browser demo search server (run_demo.py) that starts in about a minute once files land; and genuine custom-corpus tooling under open/ — distributed dumping of dense phrase vectors to hdf5, Faiss index building via run_index.py, paragraph-level TF-IDF generation with dump_tfidf.py, and the demo server itself. Training code covers SQuAD v1.1, a negative-sample finetuning pass, and a phrase classifier stage. The README is honest about scale: a Wikipedia-sized index wants a large --num_clusters (e.g. 1M) and the --hnsw flag. Where it hurts is operations. You need a Conda environment pinned to python=3.6 because Faiss 1.5.2 cannot be installed with pip, DrQA installed manually first (typically requiring java-jdk via conda), and two separate requirements.txt files — the root one for the PyTorch question-encoding server, open/requirements.txt for the search server and demo. Model, BERT, Wikipedia TF-IDF and dump files come from the gs://denspi bucket, and the dump alone is 1.5 TB. Hardware floor for serving is 4 CPU cores, 30 GB RAM and 1.5 TB SSD with no GPU; on Google Cloud the authors used 8 vCPUs, 30 GB RAM and 6 local SSDs and recommend local SSD for cheaper, lower-latency disk access at the cost of persistency. Training your own model is the expensive part: 4x P40 (24 GB) GPUs for the specified batch size, roughly 16 hours for three epochs on SQuAD v1.1, a further finetuning pass with negative samples, then under four hours for the phrase classifier. Answers are tied to a fixed Wikipedia snapshot and accuracy drops on out-of-domain questions. There is no conversational or multi-turn layer and no vendor support channel. Treat Denspi as a research artifact and a benchmark, not a product.