3LC.AI的动态

Instead of finetuning your LLMs, try dynamic few-shot prompting instead ?? With dynamic few-shot prompting, instead of injecting a fixed set of examples into the prompt, you retrieve a dynamic set of examples based on the query - so you find relevant examples that are relevant towards solving your input task. This is helpful for use cases like customer support, text-to-SQL, structured output, and more. RS Rohan has a great resource repo showing how this works using LlamaIndex workflows, check it out: https://lnkd.in/g58qxa8X For more details on workflows: https://lnkd.in/giseEZ5q

Imagine you save a bunch of your conversations with a model, and label each response as good or bad, and do rag on it. Now, when you do another question you use rag to find "good answer examples" and "bad answer examples", and pass them to the model. In this way the model will "learn with you" as time goes by. Really the technology we have at hand is a building block for more and more to come, we just need to build our own castles.

Instead of finetuning your LLMs, try dynamic few-shot prompting instead

Cost effective and worth trying

Take a look at the next in the series of blogs on what we are building at Digital Mirror. In this second post, I start to give some small insights into the core new process and framework we are building. That are even more important now when using Agents or Automation. In future posts I will expand further.

There are different ways to classify business rules. Depending on the point of view we take, we will model them differently in code which will impact readability, testability and changeability. One way to classify is whether the rule depends on the state of the system: - Email must be in valid format (does not depend on state) - Email must be unique (does depend on state) We might be tempted to group these in an email validation component because they are both rules about email. But they are not the same type of rule and serve different purposes. Rules that don't depend on state are often good candidates for value objects. #ddd #domainDrivenDesign

In last 2 weeks I was asked by the customers several times, how to improve the quality of GenAI output. My top 2 are: 1) using parallel alternative generation and summarize after 2) instead of free generation request use "choose from available options" when applicable. Here is one more example of option 1 ! #genai #llm #aiinbusiness #aicoe

Not every notification needs to delivered immediately. Sometimes, it’s about delivering the right action at the perfect time—whether that’s during specific hours, on certain days, or after batching critical data. In modern workflows, timing isn’t just important—it’s everything. With Novu’s digest engine (batch processor), you can control exactly when users progress to the next step in their workflow journey, ensuring every action happens at the right moment. Digest Action allows you to define precise periods for user progression. When users trigger an action outside this window, Novu’s Digest activates a timer, ensuring progression aligns with your predefined schedule. How It Works: 1. Define a time window: Set allowed days and hours for progression (e.g., Tuesdays, 1 PM–6 PM PST) 2. Queue out-of-window actions: If triggered outside the window, the timer holds the action until the next eligible timeframe. 3. Optimize timing: Actions are processed at strategic moments—like 5:45 PM on the next available Tuesday Example: Imagine notifying users only during weekends in the evening. With Novu, you can set a time window to ensure users progress to the message node only during this period. Any triggers outside the window are intelligently queued and processed within the defined schedule. This isn’t just about scheduling—it’s about workflow intelligence. By leveraging Novu’s Digest Engine you create workflows that are intentional, scalable, and aligned with user behavior. What are your thoughts on time window handling?

?? What are capability statements and what are they used for? Join us in this #FREE #webinar where we will present the who, what, when, where and why of capability statements: https://bit.ly/3AxCqWc

A large part of what we do involves working with disparate knowledge sources to build the best response possible to a customer query based on the intent identified. I’ve connected this to my inbox, which serves as a good example of what can be produced when combining Intent + Knowledge + Real-time information - calendar events - in this example. Tone and structure of a response are no longer an issue with the latest models adhering exactly to the structure you engineer into a prompt. The only woodenness here is due to the examples I’ve provided in the prompt. Surprisingly difficult to create some natural ones when coming up with examples...drop me a message if you would like to learn more about what we can do to automate your support.