DSPy
Use DSPy when you can define a scoring function and examples before optimising a language-model pipeline.
- Starting price
- MIT software; inference extra
- Free access
- Library free; model calls separate
- Billing unit
- Your model tokens and optimisation runs
- Platform
- Python package in your environment
- Verified
- 29 Oct 2026
Overview
What you provide
- Define task signatures, representative examples, an approved model connection and a metric that reflects useful outputs.
What you get
- DSPy returns structured predictions and reusable optimized program configurations rather than a hosted finished application.
Choosing this tool
DSPy is useful for extraction, retrieval or agent programs whose behaviour should improve against a measurable task. Typed signatures declare inputs and outputs; modules choose execution strategies and optimizers tune prompts against examples. Pydantic AI is a closer choice for typed tools and an application agent loop, while LangChain offers a wider integration-oriented orchestration stack. Optimizing against a weak or leaked metric can improve a score without improving real task quality.
A decision worked through
If an optimization evaluates fifty candidates on twenty examples, that can involve roughly one thousand example evaluations before retries or a more complex module’s extra calls. Estimate the actual call graph and provider rates; a free library does not make that experiment free.
DSPy alternatives for metric-driven optimisation of llm programs
Compare DSPy with PydanticAI, LangChain by the actual task, billing unit and operating responsibility.
| Product | Price / unit | Free allowance | Input → output | Key limit |
|---|---|---|---|---|
| DSPyRead profileOfficial source ↗ | MIT software; inference extra | No bundled model-call quota | Examples and metrics produce optimized programs. | Optimization can multiply model calls. |
| PydanticAIRead profileOfficial source ↗ | Open-source SDK; provider fees separate | Offline test model available | Typed requests become validated Python results. | Durability needs a configured backend. |
| LangChainRead profileOfficial source ↗ | MIT library; models and hosted services extra | Open-source package without inference credits | Models, tools and data form application workflows. | Provider adapters do not unify every model capability. |
Setup & workflow
- Implement a minimal signature and test it on a few known examples.
- Write a metric and split training examples from the held-out set.
- Run a bounded optimization and compare unseen failures before deployment.
Requirements & installation
- Use Python 3.10+ and install the package, then configure a model or local endpoint and bounded evaluation data.
Limits & review
- Optimizers consume model calls and can overfit examples. Typed output structure does not establish factual correctness, and external tools retain their own permissions and charges.
Your part
- Keep a held-out evaluation set and inspect failure cases before deploying an optimized program.
Sources & verification5
This profile is based on official sources, not a hands-on product test.
Vendor descriptions and demos document advertised features. Editorial guidance is based on these sources.
- Official DSPy pricingdspy.ai
- Official DSPy productdspy.ai
- DSPy official referencedspy.ai
- PydanticAI official referencepydantic.dev
- LangChain official referencegithub.com
