LLM Configuration
Artemis calls language models directly for Hyperion, the Atlas agent,
Deimos, Course Home welcomes and the embeddings of global search. All of them read one
configuration block, artemis.llm. Iris and Athena are separate services; configure their models in Pyris and Athena.
The configuration has three parts:
- Endpoints say where a provider is and how Artemis signs in to it.
- Presets name one model on one endpoint, with its options.
- Assignments in each module say which preset each task uses.
Artemis has no built-in models, endpoints or prices. A module that is enabled and has no assignment stops the server at startup, and the error lists every missing key at once, ready to copy.
A complete example
artemis:
llm:
endpoints:
logos: # reports the USD cost of each call, so it needs no price list
base-url: https://logos.example.org/v1
api-key: ${LOGOS_API_KEY}
openai: # reports no cost: Artemis estimates it from the price list
base-url: https://api.openai.com/v1
api-key: ${OPENAI_API_KEY}
prices-usd-per-million:
- { model: gpt-6.1-terra, input: 1.00, cached-input: 0.05, cache-write: 1.25, output: 5.00 }
- { model: gpt-6-luna, input: 0.10, cached-input: 0.01, cache-write: 0.125, output: 0.50 }
presets:
large: { endpoint: logos, model: gpt-6.1-terra, api: responses, reasoning-effort: medium }
small: { endpoint: logos, model: gpt-6-luna, api: responses, reasoning-effort: low }
atlaschat: { endpoint: logos, model: gpt-5.6-terra, api: chat-completions, reasoning-effort: none }
selfhosted: { endpoint: logos, model: Qwen/Qwen3.8-27B, api: chat-completions, max-tokens: 8000 }
embed: { endpoint: logos, model: Qwen/Qwen3-Embedding-8B, api: embeddings }
generation: { endpoint: logos, model: gpt-6.1-terra, api: responses, reasoning-effort: medium, max-retries: 0 }
generationthorough: { endpoint: logos, model: gpt-6.1-terra, api: responses, reasoning-effort: xhigh, max-retries: 0 }
atlas:
atlasllm:
enabled: true
llm-presets: { default: small, orchestrator: large, chat: atlaschat }
hyperion:
enabled: true
llm-presets:
default: large
problem-statement-rewrite: small
faq-rewrite: small
metadata-suggestion: small
competency-context: small
quiz-pool:
generator-llm-preset: selfhosted
judge-llm-preset: selfhosted
exercise-generation:
enabled: true
agent:
llm-preset: generation
default-profile: standard
profiles:
standard: { llm-preset: generation, max-turns: 60, max-tokens-per-job: 3000000 }
thorough: { llm-preset: generationthorough, max-turns: 100, max-tokens-per-job: 6000000 }
deimos:
enabled: true
llm-preset: selfhosted
weaviate:
enabled: true
vectorizer-module: text2vec-openai
embedding-llm-preset: embed
Endpoints
An endpoint is one provider address with its credentials. Several presets can share one endpoint; Artemis keeps one HTTP client per endpoint.
| Key | Default | Meaning |
|---|---|---|
type | openai | openai for OpenAI and every OpenAI-compatible server or gateway, microsoft-foundry for Azure OpenAI / Microsoft Foundry. |
base-url | required | The API base URL, without credentials, query or fragment. For OpenAI-compatible servers it ends with /v1, for example https://api.openai.com/v1 or http://localhost:11434/v1. |
api-key | none | Sent as a bearer token (openai) or as the Azure key header (microsoft-foundry). Without a key, Artemis sends no Authorization header. |
timeout | 300s | Time limit of one request. |
max-retries | 3 | Retries of the provider SDK after a failed request. |
organization-id | none | OpenAI organization, if your account needs one. |
microsoft-foundry-service-version | none | Azure API version, only for microsoft-foundry. |
custom-headers | none | Extra HTTP headers as a map, for example for a gateway. |
prices-usd-per-million | none | Prices for cost estimates, see Costs. |
For microsoft-foundry, the api-key is required. The model of a preset on this endpoint is the name of the
deployment. A base-url that ends with /openai/v1 uses the unified Foundry API; a resource URL such as
https://<resource>.openai.azure.com uses the deployment paths.
Presets
A preset is one model on one endpoint, plus the options that Artemis sends with every call.
| Key | Meaning |
|---|---|
endpoint | Required. The name of an endpoint. |
model | Required. The model name the provider expects, for Microsoft Foundry the deployment name. |
api | Required. responses, chat-completions or embeddings. Use responses when the provider supports it; reasoning models need it for tool calls. |
reasoning-effort | For example none, low, medium, high, xhigh, as the model supports. |
temperature | Only without reasoning-effort or with reasoning-effort: none. |
top-p | Nucleus sampling, as the model supports. |
verbosity | Text verbosity, as the model supports. |
max-tokens | Output token limit. Set at most one of max-tokens and max-completion-tokens. |
max-completion-tokens | Output token limit for models that need this field name. |
timeout | Overrides the endpoint's timeout for this preset. |
max-retries | Overrides the endpoint's max-retries for this preset. |
An embeddings preset needs an endpoint of type openai and takes none of the model options, timeout or
max-retries. Embedding calls also ignore the endpoint's timeout and max-retries: their time limits are set under
artemis.weaviate.embedding, and failed index writes are retried by the index itself.
Names of endpoints and presets use only lowercase letters and digits, and start with a letter, for example large or
gpt6terra. Environment variables cannot carry dashes or dots inside a name.
Assignments
Each module assigns presets in its own configuration block. A single task uses an llm-preset key. A module with
several tasks has an llm-presets map: each task can have its own key, and a task without one uses default. The
default key works the same in every map.
| Module and condition | Keys |
|---|---|
Hyperion, artemis.hyperion.enabled | artemis.hyperion.llm-presets with default, problem-statement, problem-statement-rewrite, consistency-check, checklist, assessment-criteria, quiz-generation, code-generation, variants, faq-rewrite, metadata-suggestion, competency-context |
| Hyperion quiz question pool, optional | artemis.hyperion.quiz-pool.generator-llm-preset and artemis.hyperion.quiz-pool.judge-llm-preset: both or neither. Without them the pool is off. |
Hyperion exercise generation, artemis.hyperion.exercise-generation.enabled | artemis.hyperion.agent.llm-preset, and optionally artemis.hyperion.agent.profiles.<name>.llm-preset. These presets need max-retries: 0. |
Atlas agent, artemis.atlas.atlasllm.enabled | artemis.atlas.atlasllm.llm-presets with default, chat, orchestrator, worker, flavor-strip |
Deimos, artemis.deimos.enabled | artemis.deimos.llm-preset |
Course Home, artemis.course-home.ai.enabled | artemis.course-home.ai.llm-preset, with location (LOCAL or CLOUD) for the students' consent, see Course Home AI |
Global search, artemis.weaviate.vectorizer-module: text2vec-openai | artemis.weaviate.embedding-llm-preset, a preset with api: embeddings |
The quiz question pool and Deimos send lecture material or student code to their model. Choose a preset on an endpoint that your institution allows for that data, for example a self-hosted model.
Exercise generation retries failed model calls itself and counts the cost of every call. SDK retries would send
requests that Artemis cannot count, so the server refuses an exercise generation preset whose effective
max-retries is not 0.
AI Worker nodes
AI Worker nodes run exercise generation. Core sends the name of the preset to the worker, and the worker looks it up in
its own artemis.llm block. Give core and every worker the same artemis.llm block. A worker without the preset, or
with retries for it, fails the run at its start. Upgrade core and workers together: worker protocol version 6 carries
the preset name.
Costs
Artemis shows LLM costs in USD in token usage tracking. For each call it uses, in this order:
- The USD amount that the provider reports with the response, for example from Logos. A reported zero is a valid cost.
- Otherwise, the entry of
prices-usd-per-millionon the preset's endpoint whosemodelis exactly the preset'smodel. - Otherwise, the call stays unpriced.
The price list is matched by the model name in the preset, not by the name in the provider's answer, so dated snapshot
names such as gpt-6.1-terra-2026-10-01 need no own entry. Each entry needs model, input and output;
cached-input and cache-write are optional. All prices are USD per million tokens. At startup Artemis warns about a
preset whose endpoint has a price list without an entry for the preset's model.
Environment variables
Every key can also be set as an environment variable. Write the key in upper case, replace dots with underscores and remove the dashes:
ARTEMIS_LLM_ENDPOINTS_LOGOS_BASEURL=https://logos.example.org/v1
ARTEMIS_LLM_ENDPOINTS_LOGOS_APIKEY=...
ARTEMIS_LLM_PRESETS_LARGE_ENDPOINT=logos
ARTEMIS_LLM_PRESETS_LARGE_MODEL=gpt-6.1-terra
ARTEMIS_LLM_PRESETS_LARGE_API=responses
ARTEMIS_LLM_PRESETS_LARGE_REASONINGEFFORT=medium
ARTEMIS_HYPERION_LLMPRESETS_DEFAULT=large
ARTEMIS_HYPERION_LLMPRESETS_FAQREWRITE=small
ARTEMIS_ATLAS_ATLASLLM_LLMPRESETS_DEFAULT=large
ARTEMIS_DEIMOS_LLMPRESET=large
Spring binds environment variables less strictly than YAML. To catch typing errors, Artemis refuses an
llm-presets variable that names no task of its map. Set a whole prices-usd-per-million list in one place, either in
YAML or in environment variables, because a list from one source replaces a list from another.
Check the configuration
At startup Artemis checks the endpoints, the presets and the assignments of the modules that run on the node. The
server stops when it finds a problem and lists all problems in one message. When the configuration is valid, Artemis
logs one line per assignment with the preset, model, API, host, timeout, retries and authentication. It adds no fallback price when the price list has no entry for the model:
LLM artemis.hyperion.llm-presets.faq-rewrite -> small | gpt-6-luna | responses | logos.example.org | timeout PT5M, retries 3, auth key, no fallback price
Upgrade from the previous configuration
Earlier releases configured models under spring.ai.openai, in module keys and with built-in defaults. Artemis now
reads none of these keys. It logs a warning that lists the old keys it finds, and stops at startup if an enabled
module has no new assignment.
| Previous key | Now |
|---|---|
spring.ai.openai.base-url, api-key, microsoft-foundry, timeout, max-retries | An endpoint under artemis.llm.endpoints |
spring.ai.openai.chat.model and its options | A preset under artemis.llm.presets, assigned with artemis.hyperion.llm-presets.default |
artemis.llm.model-costs | prices-usd-per-million on the endpoint |
artemis.atlas.chat-model, temperature, flavor-strip-*, artemis.atlas.orchestrator.model, worker-model, reasoning-effort, worker-reasoning-effort, responses-api-enabled | artemis.atlas.atlasllm.llm-presets; the preset's api replaces responses-api-enabled |
artemis.hyperion.quiz-pool.base-url, api-key, allowed-models, generator-model, judge-model, timeout | artemis.hyperion.quiz-pool.generator-llm-preset and judge-llm-preset |
Model options of artemis.hyperion.agent.profiles.<name> | artemis.hyperion.agent.profiles.<name>.llm-preset |
artemis.deimos.llm.* | artemis.deimos.llm-preset |
artemis.course-home.ai.base-url, api-key, model | artemis.course-home.ai.llm-preset |
artemis.weaviate.open-ai-base-url, gpu-api-key, open-ai-embedding-model | An embeddings preset, assigned with artemis.weaviate.embedding-llm-preset. The endpoint's base-url now ends with /v1 |