Mu Sun

Hey Matt! DSPy is a declarative programming model. It's sort of like relational databases / SQL. You write what you actually want to say (and want others to read) in English, knowing that it may or may not be the best "prompt" for your LLM. DSPy then gives you the tools to translate that clean English + code/structure + evals into the form that your LLM needs to perform best. This can take three forms: 1. Optimized prompts, often very specific to your LLM and your evals (e.g., tricks like specific few-shot examples that work best or very specific ways of repeating certain instructions, or ALL CAPS or whatever nonsense you don't want to maintain in your code). 2. RL updates to your LLM weights. You certainly would do this by hand, but DSPy can automate RL for your pipeline *exactly* with the same API as prompt optimization. 3. Inference scaling, like applying techniques that make more extensive use of your LLM(s) at inference time. Let's focus on #1 here. When you iterate with evals by hand to "prompt engineer", you're doing two things. A) You're clarifying your own specs and correcting important details. This is amazing and crucial. It's NOT handled by DSPy in the general case; you're still supposed to iterate on programming your system. B) You're overfitting to a specific LLM's nuances and failure modes, and making a lot of low-level choices to appease that model until it works. You're moving instructions to different parts of the prompt, asking for XML instead of JSON, repeating yourself for a few key things, asking the model to pretend it's Einstein, etc. These "tricks" aren't bad; sometimes they work. Sometimes they're even essential. But you shouldn't be the one hard-coding them into your application. Tricks are best handled by a compiler that has global view of your system and that builds a quality model of your LLM, so it can iterate on your behalf. When you end up switching models or sharing your code with others, this all proves invaluable. When you decide you want to RL, this all proves invaluable. When the LLM(s) get so good at your task such that hacks are not necessary, this all proves invaluable. When you want to look at pleasant clean code, not weird formatting/parsing tricks, a few months later, this all proves invaluable. Hope this helps!

我们结合会员群经常提问讨论的话题和出海孵化的实际经验， 整理了一份专门给「首次尝试」 出海产品小伙伴的 2025独立 开发出海扫盲手册（老手请自动绕行） 帮助梳理从构思、产品验证到推广运营的关键步骤，让你少走弯路、更高效地将创意变为赚钱工具。 📌 完全免费，领取方式： • 关注 @chuhaiqu • 点赞 转发并评论：出海去 会通过推特私信自动（三者缺一不可，是触发自动发送的必要条件）发送相关资料。

curl docs.jina.ai This is our Meta-Prompt. It allows LLMs to understand our Reader, Embeddings, Reranker, and Classifier APIs for improved codegen. Using the meta-prompt is straightforward. Just copy the prompt into your preferred LLM interface like ChatGPT, Claude, or whatever works for you, add your instructions, and you're set. In this example, we copied the entire prompt into Anthropic Claude and asked it to grab every sentence from Hacker News front page and visualize them using UMAP with matplotlib. This task is nontrivial as it combines multiple APIs from our Search Foundation, like Reader and Embedding where Claude may not have knowledge of. So if you asked Claude directly, it probably wouldn't give an optimal answer. But with the meta-prompt, Claude now has good knowledge about our APIs and can generate much better code! We can copy paste the code directly to Google Colab and with minimum modification, the code just works!