Selecting the ideal Indexing API for your application can significantly impact the effectiveness of integrating Vectara’s search functionalities into your application. The best indexing method depends on your needs, such as when you have semi-structured or unstructured documents, or if you want more granular control over the data segmentation and indexing process.
Vectara offers the following indexing APIs for these different scenarios:
If you want to extract text from existing, unstructured documents in common file types with minimal manual intervention, use the File Upload API. This option enables you to attach additional, user-defined metadata at the document level.
You can also upload JSON versions of the same Document protocol buffers passed to the standard indexing API and the low-level indexing API, as long as the file ends with the
.jsonextension. Our platform intelligently determines which flavor of document proto it's looking at. Note that sending any other kind of JSON to the indexing endpoint will cause it to error out.
We recommend this option if you have not written your own extraction logic already.
If you have structured documents that you want Vectara to index and segment into chunks for you, use the standard indexing API. In Vectara, a
documentis very flexible in what it can represent. It can be as short as a tweet or as long as the 1600 page Bible. The
documentobject typically includes unique identifiers like
metadatathat you can leverage. The document is also broken down into sections. Each
sectioncan have a unique
metadata. Each section can also contain other sections.
We recommend this option for applications where documents already have a clear and consistent structure like news articles, product descriptions, rows in database tables or CSV files, or records from an ERP system.
For the most advanced use cases, if you want full, granular control to chunk your document into
parts, use the low-level indexing API. These documents also have a unique ID and metadata, but you also define individual document
partswhich make up granular sections of the overall document container. These parts define the actual text to be indexed. Each part is converted into exactly one vector in the underlying index. Each part can contain individual
metadata, as well as custom dimension values that affect ranking results.
We recommend this option for Machine Learning teams with expertise in neural information retrieval who want low-level control over how documents are indexed in our systems. Using the low-level API typically involves significant coordination between your Machine Learning team and organizational stakeholders.