> ML_LIBRARY // DSPY_v1.0
DSPy
Stanford NLP Group / Omar Khattab — Programming—not prompting—Foundation Models.
nlp-llmv2.5.0MITqualified
Model Training
Accelerators:
CPUCUDA
Distributed Training:No
Model Inference
Inference Accelerators:
CPUCUDA
Deployment Targets:server
What It Does
- +Replaces fragile manual prompt strings with declarative Signatures and Modules (Predict, ChainOfThought)
- +Automatically optimizes prompt instructions and few-shot examples using optimizers (MIPROv2, BootstrapFewShot)
- +Compiles multi-step LLM programs against quantitative metric assertions
What It Does Not Do
- -Train neural network backpropagation weights directly (optimizes prompts and weights via LM calls)
- -Serve low-level GPU tokens (relies on underlying model APIs or vLLM)
- -Run in web browsers
>Suitable Work Types
- Replacing fragile 500-word prompt strings with maintainable, compiled software code
- RAG retrieval-augmented pipelines requiring continuous metric-driven prompt refinement
- Multi-hop question answering and automated evaluation
>Unsuitable Work Types
- Standard pre-training of foundational weights
- Classical tabular algorithms where scikit-learn is superior
Data Residency Implications
Transmits optimization iterations to configured LLM API endpoints.
Security Considerations
Optimized prompts must be inspected for policy compliance prior to production promotion.
Operational Profile & Known Limitations
Maturity:emerging
Learning Curve:high
Ops Complexity:moderate
Cost Tier:medium
> Known Limitations:
- Compiling programs with teleprompters (e.g. MIPRO) requires hundreds of LLM calls, incurring API costs during optimization.
- Requires adopting a programmatic mental model distinct from traditional string prompting.
Associated Incident Patterns (Incidentpedia)
Enforce safeguards and monitoring to guard against these documented real-world failure modes:
> Primary Evidence & Benchmark Citations
DSPy Documentationofficial-docs • >=2.4.0, <=2.5.x
2026-09-25HIGH