Markdown and HTML Responses Overview

In response to feedback from our developer community, Vectara has introduced the capability to define generative responses in a structured format. This innovation allows developers to receive search responses in markdown or HTML format, with references and metadata cleanly integrated using URLs from document metadata fields. This structured approach not only simplifies parsing on the client side but also reduces the likelihood of errors significantly.

Starting today, you can ask Vectara to return Markdown or HTML links (<a> blocks with href attributes) in its cited summary.  Those can include document metadata directly, giving you the ability to reference document titles, PDF page numbers, document IDs, URLs, and other metadata.  This works by 2 new citation formats with custom parameterization.  For example, say you wanted Vectara to say “as seen in [Document-Title]” with a link to the specific page, you can now do that by:

1. Setting the style of the citationParams to either MARKDOWN or HTML
2. Setting the urlPattern to e.g. {doc.id}#page={section.page}
3. Setting the textPattern to As seen in {doc.title}

All together, this looks like:

After this, Vectara will respond with something like:

You could also set style to HTML and instead of Markdown-formatted links, Vectara would respond with something like …

Now this allows you to make use of the response directly in your Markdown- or HTML-enabled user interface without any additional parsing logic!  For more information on how to use the new citation formats, see the documentation and have a look around our API playground.

Let’s get on to the next feature we’re releasing today: semantic search across conversation histories.

Semantic Conversation History Search

Everyone has users that ask “unexpected” questions.  Let’s suppose again that you’re operating an MTA chatbot and you have riders that are asking questions like whether they can bring their pet llamas or ETs onboard:

Vectara has tried to answer the question and cited its sources, but in the event these types of questions are common, you might want to add additional information to the source information so your users get more direct answers.  But how do you find these “unknown unknowns?”

With the new semantic conversation search feature released in Vectara today, you can now find instances where the information wasn’t known:

And even users that might have been upset and used arbitrary language indicating that they’re upset:

To get this behavior, we didn’t have to set up any rules that said “an upset user might say ‘where the heck’ …” or tell Vectara what an “unknown response” was: it was able to use the power of its semantic search capabilities right out of the box.

Conclusion

Vectara’s introduction of structured generative response formats and semantically searchable conversation histories mark a significant leap forward in retrieval augmented generation (RAG). By simplifying the integration process and administrator visibility, Vectara continues to empower developers and GenAI chat administrators to build more engaging and user-friendly applications and plug gaps in their knowledge bases. These innovations not only reflect Vectara’s commitment to addressing the needs of its developer community but also sets a new standard for usability and effectiveness.

As always, we’d love to hear your feedback! Connect with uson our forums oron our Discord.

If you don’t have an account already, sign up for a free Vectara account to see how Vectara can help you easily leverage RAG in your GenAI apps!

Vectara Sign-UpGet started with Vectara today!Sign Up!