KB search does not understand structured product data | Add system instruction to KB search
Hello dear community,
I am currently trying to build a product advisor as part of a shopping assistant in Voiceflow, the product data for this runs directly into the knowledge base. The relevant content is structured in both - the metadata and the chunks.
Unfortunately, when performing a search, the available information is sometimes not interpreted or searched correctly. To give a few examples:
Example 1: A user enters the following question in the search field: “I am looking for pants in size 26U.” Result: Pants are found, when I try this search in English, but when I try the same question in German "Ich suche eine Hose in Größe 26U", there is no result.
Example 2: A user enters the following question in the search field: “I am looking for an item with 46.00% Lyocell.” Result: No items are found, even though there are relevant articles in the KB (see screenshots attached).
Simple search queries such as “I am looking for a blue polo shirt for women” work fine.
I now have two questions in this context:
1. Why is the data stored in the knowledge base not being recognized and read correctly? Especially they are definitely available in the chunks and metadata.
2. Is there a way to provide the system with additional explanatory instructions for searching the knowledge base? I tried to include these instructions in a variable in the API block via “system” (see screenshot), but this does not seem to have any effect.
Example: “If the user asks for an item in a specific size, only items that are available in the requested size should be returned. The sizes in which an item is available can be found in the metadata under chunks.metadata._cf_alle_groessen.”
I am grateful for any help!
1 Reply
Hello man @Marc B., I can clearly see where u are going wrong as I had similiard issues - there is multiple things you need to take into account:
1) Your KB - your KB must be well structured in tabular form if you are dealing with products where as each product must have detailed matadatafields populated
2) Construct metadatafield filters dynamically via LLM based on the user query in order for you to narrow down the KB data set and extract only the most relevant products
3) Query optimization - You need to optimize the query into a RAG / vector search version enriched with most pertinent keywords based on the products you are looking for + the query must be in the same exact language as are the products in the KB
Shoot me DM, if you want to go more in depth 🙂