Settings¶

Settings are read from environment variables only; no file is loaded. In Python, pass LPGCastSettings(...) to LPGCaster instead.

Model¶

Variable	Default	Meaning
`LPGCAST_LLM_MODEL`	`gpt-5.6-luna`	Model name sent to the endpoint. The default is an OpenAI model; set it for any other endpoint.
`LPGCAST_LLM_BASE_URL`	unset	Endpoint base URL, e.g. `http://localhost:11434/v1` for Ollama. Unset means OpenAI.
`LPGCAST_LLM_API_KEY`	unset	API key. Required for any provider call, including to endpoints that ignore it (set any value for a local server). Without one only cached responses are served.
`LPGCAST_LLM_TEMPERATURE`	unset	Sampling temperature. Unset sends none: the model's default.
`LPGCAST_LLM_REASONING_EFFORT`	unset	Passed through to models that accept `reasoning_effort`.
`LPGCAST_LLM_MAX_OUTPUT_TOKENS`	unset	Output ceiling per call: `max_completion_tokens` to OpenAI, `max_tokens` to any other endpoint.
`LPGCAST_LLM_RESPONSE_FORMAT`	`auto`	`json_schema` sends the instance schema for strict decoding; `json_object` asks for JSON and puts the schema in the prompt; `auto` tries `json_schema` and falls back to `json_object` for the rest of the run when the endpoint rejects it.
`LPGCAST_LLM_REQUEST_TIMEOUT_SECONDS`	`120.0`	Per-request timeout.
`LPGCAST_LLM_TRANSPORT_RETRIES`	`2`	SDK retries on connection errors, 429 and 5xx.
`LPGCAST_LLM_PARSE_RETRIES`	`2`	Re-asks after a response that is not a JSON object, with the parse error shown to the model.
`LPGCAST_LLM_MAX_INFLIGHT`	`8`	Concurrent provider requests per event loop.

Variable	Default	Meaning
`LPGCAST_CACHE_ENABLED`	`true`	Read and write cached responses.
`LPGCAST_CACHE_DIR`	unset	Cache directory. Unset: `./.test_cache` under pytest, else `$XDG_CACHE_HOME/lpgcast` or `~/.cache/lpgcast`.
`LPGCAST_CACHE_READ_ONLY`	`false`	Serve cached responses but write no new ones.
`LPGCAST_CACHE_REQUIRE_HIT`	`false`	A cache miss raises instead of calling the model: a replay that provably makes no request.

Variable	Default	Meaning
`LPGCAST_MAX_EPISODE_TOKENS`	`4000`	Ceiling on an episode's estimated tokens (characters / 4). Episodes are never chunked: one over the ceiling is refused before any call.
`LPGCAST_SLICE_MAX_TOKENS`	`4000`	Token budget of the schema slice.
`LPGCAST_SLICE_MAX_ELEMENTS`	`60`	Vertex and edge budget of the schema slice.
`LPGCAST_USER_INSTRUCTION`	unset	Extra guidance placed after the rules, before the text.

Variable	Default	Meaning
`LPGCAST_SERVE_API_KEY`	unset	Bearer token every request but `/health` must carry. Without one the service refuses to cast unless `allow_anonymous` is set.
`LPGCAST_SERVE_ALLOW_ANONYMOUS`	`false`	Serve without a key; for local development only.
`LPGCAST_SERVE_MAX_EPISODES`	`200`	Most episodes one request may carry.
`LPGCAST_SERVE_HOST`	`127.0.0.1`	Address to bind.
`LPGCAST_SERVE_PORT`	`8997`	Port to bind.