Amin Ahmad — CEO, ZIR.AI (now Vectara.com) — Algolia / Elasticsearch-like search product on neural search principles
Amin Ahmad, Former NLP researcher at Google, who developed projects like Google Talk To Books, shared with me his journey in his startup called ZIR.AI (now vectara.com)— building Algolia / Elasticsearch-like search product on neural search principles.
In this episode you will learn what it takes to productionize a neural search system, including aspects of model distillation and SLA guarantees, as well as data augmentation. We also spoke about the vector search as an industry at large and Search Engineer profession transformation with neural search in the game. And of course quite a few papers were mentioned — all in show notes for your study. I’m hoping these episodes are educational, as they are fun to watch.
Be sure to check out the show notes to get a discount code for an extended trial of the ZIR.AI (now vectara.com) search platform for your documents.
Topics:
00:00 Intro
00:54 Amin’s background at Google Research and affinity to NLP and vector search field
05:28 Main focus areas of ZIR.AI in neural search
07:26 Does the company offer neural network training to clients? Other support provided with ranking and document format conversions
08:51 Usage of open source vs developing own tech
10:17 The core of ZIR.AI product
14:36 API support, communication protocols and P95/P99 SLAs, dedicated pools of encoders
17:13 Speeding up single node / single customer throughput and challenge of productionizing off the shelf models, like BERT
23:01 Distilling transformer models and why it can be out of reach of smaller companies
25:07 Techniques for data augmentation from Amin’s and Dmitry’s practice (key search team: margin loss)
30:03 Vector search algorithms used in ZIR.AI and the need for boolean logic in company’s client base
33:51 Dynamics of open source in vector search space and cloud players: Google, Amazon, Microsoft
36:03 Implementing a multilingual search with BM25 vs neural search and impact on business
38:56 Is vector search a hype similar to big data few years ago? Prediction for vector search algorithms influence relations databases
43:09 Is there a need to combine BM25 with neural search? Ideas from Amin and features offered in ZIR.AI product
51:31 Increasing the robustness of search — or simply making it to work
55:10 How will Search Engineer profession change with neural search in the game?