Imp 는 언어 모델 단계의 입력과 출력을 선언하고, 예제와 평가 지표로 프로그램을 개선하는 Elixir 프로젝트로, DSPy의 시그니처, 모듈, 최적화기, 에이전트 루프와 검색 기능을 OTP 의 신뢰성과 동시성 위에서 실행함 시그니처 로 프롬프트를 생성...

declarative self-improving language-model programs for Elixir 😇 - deepfates/imp

declarative self-improving language-model programs for Elixir 😇 - deepfates/imp

declarative self-improving language-model programs for Elixir 😇 - deepfates/imp

declarative self-improving language-model programs for Elixir 😇 - deepfates/imp

declarative self-improving language-model programs for Elixir 😇 - deepfates/imp

declarative self-improving language-model programs for Elixir 😇 - deepfates/imp

Comments

What Happened Agent frameworks and SDKs (examples include LangChain, LangGraph, LlamaIndex, AutoGen, CrewAI, PydanticAI, DSPy, Semantic Kernel, OpenAI Agents SDK and Claude Code) are converging on the same architectural patterns:…

This has an actual name in the literature, “query clarification” or “ambiguity-aware RAG,” so worth knowing the terms even if you’re experimenting freeform. Concretely to try: DSPy is the best fit for hands-on experimenting since it lets you define a “clarify or answer” module and optimize the decision boundary with real examples rather than hand-tuning prompts. For frameworks with clarification built in already, LangGraph lets you build an explicit branch (ambiguous → ask, confident → retrieve+answer) as a graph rather than a single prompt, which makes the logic easier to inspect and tweak than a monolithic RAG chain. Two research patterns worth reading before you build, since they map directly to what you’re describing: Tree of Clarification (ToC), which generates clarifying questions recursively for fuzzy queries and resolves them against the retriever rather than just asking the user everything upfront, and the simpler “Clarify Once, Learn the Default” pattern, which asks one focused question the first time a certain ambiguity shows up, then remembers the user’s typical answer so it stops asking the same thing repeatedly. That second one is probably the most practical starting point for a solo experiment, it’s a small, well-scoped loop (detect missing field → ask → store default → reuse) rather than a full research pipeline.

はじめに分類モデルの精度を上げていると、そもそも、この分類問題を解いて何がよくなるんだっけ。と思うことがある。むしろデータサイエンティストとしては、常にこれが頭になければいけない。もちろん、開発を始めるときには機械学習問題の先にある目的を説明している。問い合わせ対応を速くしたい。重大な障害を見落としたくない。担当者の負担を減らしたい。そのためにデータを集め、ラベルを付け、モデルを学習させる。ところ

Many DSPy Signatures already define decisions. Today, they can run on System One models.

Declarative prompt optimization represents a paradigm shift from heuristic-based manual prompt engineering to systematic, algorithm-driven instruction tuning for large language models.