Search and Similarity are prime functions in any CRUD application. Most e-commerce frameworks come with built-in implementations, complexity ranging anywhere between standard SQL LIKE query to generic TF/IDF.
Any production use-case at scale however needs to plug into commercial third-party solutions, quality and performance of which can vary.. services often use older models or obsolete implementations underneath with yearly contract lock-ins, or teams have to build their own systems which can be expensive, time-consuming and often out-of-reach for small teams.
We are happy to present a suite of Pay-as-you-go APIs to build a pretty good search and recommendations system in 2 API calls, our implementation and scaffolding underneath use text-embedding-3-large, flagship text embedding model by OpenAI as simple easy to configure managed search and recommendations service and bespoke implementations by combining the results with other filtering logic or engineering the keys in embedding queries.
To generate embeddings just call the endpoint with your serialized data source with relevant keys included in a list of data objects.
Don't have serialized data-source, no problem, just enter plaintext data for each object record.. don't have that either just crawl your webpages to create a state of the art embeddings using the /generate_embeddings_from_url endpoint.
Configuration is simple, and pricing is transparent and based on tokens consumed. You only pay for credits as you use, you can self-service deploy from the API docs, or we can do a 1-time configuration to build in the recommendation and search system into your preferred framework or custom deployment.
Setting up State-of-the-Art, recommendation system no longer requires enterprise deals worth thousands of dollars / mo, nor does it require hiring a developer team. You only pay as you go and scale usage as you grow.
The out-of-the-box deployment for search and recommendations engine is a two step process-
Step 1 : Generate embeddings from your data. Optionally set TTL to auto-expire data after a period of time, periodically call update_embedding to update the data in the template.
You can create embeddings in either plain text or JSON format.
If Providing embedding data in JSON format, include keys which you'd like to influence search and recommendations.
Two API endpoints are supported to create embeddings which can then be queried against or other operations-
You can create embeddings from JSON input payload by calling the /chat/create_embedding route endpoint.
Creates embeddings from a JSON input payload and saves it to a Vector DB template with provided name spaced template name.
If you simply want to create knowledge base from your website, just call the generate_embedding_from_url endpoint with relevant links.
You can also additionally chunk the JSON input or larger repositories via /chat/update_embedding
Once the embeddings have been created, you can then setup a search system to search and query against the knowledge-base.
Step 2 -
You can then POST /chat/search_against_embedding to search against the embedding list for a search term. This endpoint takes optionally a POST parameter
For Recommendations system implementation you can call the /chat/similar_to_embedding endpoint.
Both endpoints take in the input search string and optional metatadata object.
You can further augment the search results with LLM response to build a RAG based search query engine, or do location augmentation to rerank the results nearest to a given location. The possibilities are endless. The basic implementation itself gives a State of the art search and recommendation engine out of the box.
Most stores and platforms could use better search and recommendation system, however building a bespoke search and recommendations engine is usually difficult and time-consuming. Products and listings in your store probably suffer from lack of discoverability. Generic search and recommendations limit the discoverability of products even in medium sized stores and frustrate the user when they cannot find exactly what they are looking for (even though the product / experience may already exist in your store)
With our configurable APIs you get pay-as-you-go deployment for search and recommendations, similarity and many other recommendations / rankings use-cases with configurable TTLs.
Provide better recommendations and implement complex logic combining similarity and location factors. Current deployment includes complex use-cases such as "Recommended by establishment" section here for client implementation, which does a combined location ranking via embeddings and then popularity ranking and filtering within the result-set, to provide most popular nearest activities nearest to a given referring establishment.
Many bespoke recommendation + search/additional filtering criteria based implementations are possible.
Are you ready to upgrade your store experience to the next level,Drop an email at contact@samsar.one or book a 30 minute meeting here.