Industrial product search is search software for a distributor's catalogue that matches an engineer's description, or a list of specifications, to the right part. It is built for catalogues of lakhs or millions of technical items, such as bearings, valves, cables and electronic components. It is a part of the B2B ordering portal, and electronic components distributors use it too.
Ordinary keyword search fails on these catalogues for three reasons. First, buyers describe what a part does, while the catalogue lists part codes. Second, the same thing has many names: "deep groove ball bearing", "radial ball bearing" and a code such as 6210 can all describe the same item. Third, different manufacturers write specifications in different units and formats.
Industrial product search uses vector embeddings to solve this. An embedding is a list of numbers that an AI model makes from a piece of text. Texts with similar meanings get similar lists of numbers. The software makes one embedding for every product in the catalogue, from its name, description and specifications. These embeddings are stored in a vector database. When a buyer types a query, the software makes an embedding of the query too, and finds the products whose embeddings are closest to it. It then applies exact filters on numbers, such as size and temperature, because embeddings do not compare numbers well. The product data arrives cleaned and in fixed fields from product information management (PIM).
Consider a maintenance engineer at a cement plant in Chhattisgarh. A conveyor has stopped because a bearing failed, and he does not have the part code. He types: "sealed bearing for a conveyor pulley, 50 mm shaft, runs hot, around 120 degrees". The distributor's catalogue has 3 lakh items. Keyword search finds nothing useful, because no listing uses the word "pulley". The embedding search finds about 200 bearings whose descriptions are close in meaning. The exact filters then keep only those with a 50 mm bore, the hole for the shaft, and a rating of at least 120 °C. Eight bearings remain. The software shows the four that are in stock at the distributor's warehouse in Raipur, and the engineer orders two that morning.
Newer systems also use a large language model (LLM) to answer follow-up questions from the datasheet, such as a bearing's maximum speed. The answer must quote the datasheet, so that the engineer can check it before he relies on it.