LLM 0.33: Reasoning Summary and Combinable Templates

The command-line tool llm enhances its advanced uses: a reasoning_summary option arrives for response API models, templates become combinable, and embedding key management is clarified. Dependencies shift to OpenAI 3.x and httpx2. A 0.32.1 patch preceded this batch, presented as more comprehensive.

A reasoning_summary option for responses

Response API models with reasoning capabilities now accept the reasoning_summary option. Three levels are offered: auto, concise, and detailed. This option is used with the endpoint “llm openai --responses.” It is described as particularly useful for testing different models that mimic OpenAI's response API. The tracking of this development refers to ticket 1600.

Combinable and reusable templates

The command “llm prompt” now accepts the repetition of -t/--template to combine templates. It thus becomes possible to use the configuration and options of one template while utilizing the prompt of another. The mechanism is illustrated by the creation of saved aliases — for example, “lhigh” with a model and default options, or “pelican” for a given prompt — and then executing them together via “llm -t lhigh -t pelican.”

Embedding keys, compatibility, and updated dependencies

The commands “llm embed” and “llm embed-multi” now accept --key, and the embedding methods at the model and collection levels accept a key= argument per call. Using the key with each call helps avoid any modification of the shared state of the model, and plugins that access self.key remain functional thanks to backward compatibility. Furthermore, embedding models adopt the same key schema as classic LLM models. On the foundational side, the OpenAI library moves to 3.x and the HTTP client migrates from httpx to httpx2, changes associated with tickets 1608 and 1631. The addition of --key for embeddings is linked to tickets 757 and 1620. A 0.32.1 patch had preceded this delivery, presented here as more complete. Contributor ChrisJr404 is explicitly thanked.