@spec derive_family(String.t()) :: String.t() | nil

Derives the family name from a model ID using prefix logic.

Extracts family from model ID by splitting on "-" and taking all but the last segment. Returns nil if the family cannot be reasonably derived.

Examples

iex> LLMDB.Enrich.derive_family("gpt-4o-mini")
"gpt-4o"

iex> LLMDB.Enrich.derive_family("claude-3-opus")
"claude-3"

iex> LLMDB.Enrich.derive_family("gemini-1.5-pro")
"gemini-1.5"

iex> LLMDB.Enrich.derive_family("single")
nil

iex> LLMDB.Enrich.derive_family("two-parts")
"two"

@spec enrich_model(map()) :: map()

Enriches a single model map with derived and default values.

Sets the following fields if not already present:

• family: Derived from model ID
• provider_model_id: Set to model ID

Note: Capability defaults are handled automatically by Zoi schema validation.

Examples

iex> LLMDB.Enrich.enrich_model(%{id: "gpt-4o-mini", provider: :openai})
%{id: "gpt-4o-mini", provider: :openai, family: "gpt-4o", provider_model_id: "gpt-4o-mini"}

iex> LLMDB.Enrich.enrich_model(%{id: "claude-3-opus", provider: :anthropic, family: "claude-3-custom"})
%{id: "claude-3-opus", provider: :anthropic, family: "claude-3-custom", provider_model_id: "claude-3-opus"}

iex> LLMDB.Enrich.enrich_model(%{id: "model", provider: :openai, provider_model_id: "custom-id"})
%{id: "model", provider: :openai, provider_model_id: "custom-id"}

@spec enrich_models([map()]) :: [map()]

Enriches a list of model maps.

Applies enrich_model/1 to each model in the list.

Examples

iex> LLMDB.Enrich.enrich_models([
...>   %{id: "gpt-4o", provider: :openai},
...>   %{id: "claude-3-opus", provider: :anthropic}
...> ])
[
  %{id: "gpt-4o", provider: :openai, family: "gpt", provider_model_id: "gpt-4o"},
  %{id: "claude-3-opus", provider: :anthropic, family: "claude-3", provider_model_id: "claude-3-opus"}
]

@spec inherit_canonical_costs([map()]) :: [map()]

Propagates cost from canonical models to their dated variants.

For models with a date suffix (e.g., gpt-4o-mini-2024-07-18), if the model has no cost, looks up the canonical model (e.g., gpt-4o-mini) from the same provider and copies its cost.

Models that already have a cost are left unchanged.

Examples

iex> models = [
...>   %{id: "gpt-4o-mini", provider: :openai, cost: %{input: 0.15, output: 0.6}},
...>   %{id: "gpt-4o-mini-2024-07-18", provider: :openai}
...> ]
iex> [_, dated] = LLMDB.Enrich.inherit_canonical_costs(models)
iex> dated.cost
%{input: 0.15, output: 0.6}

@spec inherit_hosted_claude_reasoning([map()]) :: [map()]

Copies native Claude reasoning metadata to hosted Claude models that lack it.

Amazon Bedrock, Google Vertex AI and Azure serve Claude with the Anthropic Messages request body, so a hosted entry accepts the same thinking types and effort levels as the native model named by its canonical_model_id. Upstream sources usually publish only reasoning.enabled for hosted entries.

For a model hosted on Bedrock or Vertex AI, or one whose wire_protocol is anthropic_messages, that links to an anthropic/ canonical model, this fills capabilities.reasoning.thinking, capabilities.reasoning.effort and extra.thinking_constraints from the native model when the hosted entry has none. Values the hosted entry already sets are kept, and entries whose reasoning is explicitly disabled are left alone.

Routers that translate Claude into another request format, such as OpenRouter, are not changed.

Examples

iex> models = [
...>   %{
...>     id: "claude-sonnet-5-5",
...>     provider: :anthropic,
...>     capabilities: %{reasoning: %{enabled: true, thinking: %{types: ["adaptive", "between_tools"]}}}
...>   },
...>   %{
...>     id: "global.anthropic.claude-sonnet-5-5",
...>     provider: :amazon_bedrock,
...>     capabilities: %{reasoning: %{enabled: true}},
...>     extra: %{canonical_model_id: "anthropic/claude-sonnet-5-5"}
...>   }
...> ]
iex> [_native, hosted] = LLMDB.Enrich.inherit_hosted_claude_reasoning(models)
iex> hosted.capabilities.reasoning.thinking.types
["adaptive", "between_tools"]