DSPy 3.3.1

Apply the Python 3 baseline first. These rules cover the dspy 3.3.1 distribution on Python 3.10–3.14, including models, adapters, tools, ReAct, experimental ReActV2, RLM, optimization, and persistence.

Model configuration and input boundaries

• Construct dspy.LM with server-selected providers, models, endpoints, headers, and credentials; give distinct tenants or workloads separate provider permissions and quotas where their trust differs.
• Set process defaults with dspy.configure before serving requests and use dspy.context for trusted request-scoped overrides. Context overrides do not clone shared LM, callback, or trace objects; do not mutate those objects across tenants.
• Map request data to an explicit allowlist of business inputs before calling Predict; keep its lm, signature, demos, config, prediction, and _trace overrides under application control instead of unpacking arbitrary request dictionaries.
• Validate required inputs, types, lengths, and unknown fields before Predict. Its missing-input and type-mismatch warnings are not rejection, and warn_on_type_mismatch=True does not enforce a schema.
• Set dspy.LM output limits with max_tokens, bound num_retries, and configure provider timeouts. Budget the complete module graph, parallel workers, adapter retries, and tool calls; track_usage=True reports usage but does not enforce a spending ceiling.

Predictions, retrieval, and media

• Keep signatures and module definitions in trusted code. Treat demonstrations, retrieved passages, History, tool results, and model reasoning