Markprompt

I was speaking to someone at a unicorn company who interacts with 200 customers in a week. This is more than what any product manager at the company speaks to in a year. That person is in support. Think about how much knowledge is lost by having support removed from product. When support reports to product, magic happens: support directly contributes to improving the product, incidentally leading to… less support!

For decades, software buyers have faced a tradeoff: build or buy? But it’s now clear that buyers have a third alternative: building in close collaboration with a vendor. Buyers of all solutions will find themselves in a PayPal vs. Stripe dilemma during the procurement process: Whereas PayPal offers “one size fits all” solutions for receiving payments, Stripe takes the “infrastructure approach” of empowering product and engineering teams by doing the heavy lifting of financial transactions: eng/prod teams need not worry about transaction atomicity, fraud detection, disputes etc. - but the end user experience remains entirely under their control. So after observing this trend for years, our guiding principle at Markprompt is “Build, with us.” And to a large extent, we’re betting that this is the future of software. Companies are tasking eng teams to build bespoke gen AI solutions deeply integrated into the core product experience BUT these same eng teams often struggle to “get it right” - whether lack of expertise “taming” the LLMs, Pandora’s box of tooling to build in order to operate efficiently, measure performance, generate new knowledge, and so on. So the future of many industries looks like Stripe: keep full control of the customer experience you want to build, but build on top of a partner who’s done the heavy-lifting of data preprocessing, prompt optimization, retrieval strategies, security, data flywheels, etc.

At Markprompt, we are building AI for customer support. One of the things we see consistently is “knowledge base debt” - it’s like legacy code, but for customer support materials. At a company we recently started working with, we have seen the knowledge base reduced from 11000+ articles to under 6000 within a span of 3 months. The docs and engineering teams are now pushing updates dozens of times a day, and their docs went from calcified to vibrant and always up to date. Why? Because having an AI powered support agent analyze conversations and content can quickly uproot outdated content and inaccurate articles. Support teams receive fewer tickets from customers asking about obsolete or incorrect support content, and customers are satisfied because they can find the correct answers on their own, with no inconsistencies. “Knowledge base debt” will be critical for companies to solve: it means fewer headaches for support teams, and happier customers.

Here is what we shipped in the past weeks at Markprompt: - Automatic message processing with: - Taxonomy/triaging - Sentiment analysis - Urgency level - CSAT scores - New insights charts: - Escalation rate - Answered/unanswered rate - Citations - Support ticket deflection template with follow-up questions: markprompt.com/templates - Attach case number to a thread via our Threads API: https://lnkd.in/ezZFgsiv - Zendesk case creation integration - Slack integration - Raycast extension

A key metric for AI-enabled customer support teams is: How many support cases did the AI prevent from being created by solving the customer's problem? The new Markprompt insights brings clarity to that question. Built with the excellent Airbnb visx and shadcn/ui.

Here is what we shipped at Markprompt in January: 1. New sources: Salesforce Case and Zendesk Tickets You can now generate new answers and drafts based on past resolutions. The LLM prioritizes recent tickets over earlier ones. PII is removed at import time. 2. New RAG pipeline We have drastically improved the retrieval algorithm, ensuring that the most comprehensive and precise context is brought in to generate answers. For instance, instead of pulling in individual sections from a knowledge base, related sections are also pulled in. The algorithm adjusts its behavior depending on the model being used, so that it can fully leverage the available context window, such as the 128k tokens of GPT-4 Turbo. 3. Chat completion filters Want to only pull in certain parts of your indexed knowledge base for certain interactions? For instance, for a user on the Enterprise plan, you may want to include all articles with the “Enterprise” tag, but these should not be included for users on the free plan. With filters, you can specify rules on article metadata for what content should be used in a given situation. 4. Support for gpt-4-turbo-preview Instead of specifying a pinned version of GPT-4 Turbo, such as gpt-4-0125-preview, you can now use gpt-4-turbo-preview and always be on the latest version. Incidentally, the latest version of GPT-4 Turbo improves handling of “laziness”, where the model doesn’t always complete a task, such as a step-by-step action. 5. Improved detection of “No response” situations When the LLM cannot answer a question, it tells so in the most adapted way, using formulations that fit into the flow of the conversation. It doesn’t just leave an empty answer or say “No response”, which sometimes makes it tricky to detect. Our new LLM-based approach captures these situations robustly. 6. Advanced website importer Some websites are notoriously difficult to import, because of server settings, long loading times, dynamically rendered content, and more. Our new website importer is built for situations where a standard approach falls short. 7. Mobile widget The Markprompt chat widget is now fully responsive, and works well on devices of all sizes. 8. Sticky chat The widget also supports a new "sticky" flag, so that it remains on-screen when the user is browsing a page, or navigating to other pages. 9. Updated /threads API The threads API now offers an "expand" flag that allows you to pull in all message data in a single call.