lm15 cookbook¶

Eighteen recipes. Every code block runs; every output block is a real capture. Style rules live in STYLE.md.

Essentials¶

01 — Your first request — One model string, one Request, one response: the router as the front door, keys included.
02 — Multi-turn conversations — Conversations are tuples you build yourself; nothing is hidden in a session object.
03 — System & developer prompts — Where instructions go and how each provider maps them on the wire.
04 — Controlling generation — Config: temperature, max tokens, stop sequences, and what each provider ignores.
05 — Streaming — Typed delta events from router.stream(), sync and async, without parsing SSE yourself.

06 — Function tools: define & dispatch — tool(fn) derives the schema; you run the loop. derive_tool() explains it.
07 — Built-in provider tools — Web search, code execution, and friends: server-side tools you enable, not implement.
08 — Structured output — JSON schemas in, parsed objects out, across all three providers' strict modes.

09 — Images, PDFs & documents — Sending bytes and URLs as parts; what each provider accepts.
10 — Audio, video & reasoning models — The expensive modalities, plus thinking deltas from reasoning models.

11 — Prompt caching — Cache breakpoints that cut cost on long prefixes, with real usage numbers before and after.
12 — Batch jobs — Submit hundreds of requests as a ticket; collect the answers later at about half price.
13 — Media generation (images & speech) — Images and speech out of the same Part types you send in.
14 — Video generation — Job-shaped on every wire: submit a ticket, wait, collect the MP4.
15 — Live sessions (realtime) — Bidirectional sessions over websockets: events in, audio and text out.

16 — Local & OpenAI-compatible servers — Ollama, vLLM, and friends via OpenAIChatLM with base_url and compat.
17 — Errors, retries & testing — The typed error tree, why retry policy is yours, and testing without a network.
18 — Provider passthrough — Reaching provider-only knobs without forking your code off lm15's types.
20 — Judgments — Declared answers (choice, ordered levels, yes/no) with a probability per option, from Jev natively and from other models honestly.
19 — Prompt caching — One line that reuses a long beginning on every provider, and shows what it cost.