@spec compile_filters(
  allow :: :all | map(),
  deny :: map(),
  known_providers :: [atom()] | nil
) ::
  {%{allow: :all | :none | map(), deny: map()}, [{:unknown, [term()]}]}

Compiles allow/deny filter patterns to regexes for performance.

Parameters

• allow - :all or %{provider_atom => [pattern_strings]}
• deny - %{provider_atom => [pattern_strings]}
• known_providers - Optional list of known provider atoms for validation (defaults to all existing atoms)

Patterns support glob syntax with * wildcards via LLMDB.Merge.compile_pattern/1.

Provider keys that don't correspond to existing atoms are silently ignored.

Deny patterns always win over allow patterns.

Returns

{%{allow: compiled_patterns, deny: compiled_patterns}, unknown_providers}

Where allow patterns are :all, :none, or %{provider => :all | [%Regex{}]}, and unknown_providers is a list of provider keys that were ignored.

@spec get() :: map()

Returns normalized configuration map from Application environment.

Reads :llm_db application config and normalizes with defaults.

Configuration Format

config :llm_db,
  allow: :all,  # or [:openai, :anthropic] or %{openai: ["gpt-4*"]}
  deny: %{},    # or [:provider] or %{provider: ["pattern"]}
  prefer: [:openai, :anthropic],
  custom: %{
    vllm: [
      name: "Local vLLM Provider",
      base_url: "http://localhost:8000/v1",
      models: %{
        "llama-3" => %{capabilities: %{chat: true}},
        "mistral-7b" => %{capabilities: %{chat: true, tools: %{enabled: true}}}
      }
    ],
    custom_provider: [
      name: "My Custom Provider",
      models: %{
        "model-1" => %{capabilities: %{chat: true}}
      }
    ]
  }

Provider keys can be atoms or strings. Patterns support glob syntax with * wildcards.

Custom providers are defined with provider ID as key, and a keyword list containing:

• :name - Provider name (optional)
• :base_url - Base URL for API (optional)
• :models - Map of model ID to model config

Returns

A map with keys:

• :compile_embed - Whether to compile-time embed snapshot (default: false)
• :snapshot_source - Snapshot source descriptor (default: :packaged)
• :allow - Allow patterns (:all or %{provider => [patterns]})
• :deny - Deny patterns (%{provider => [patterns]})
• :prefer - List of preferred provider atoms
• :custom - Custom providers map (provider_id => provider_config)

@spec sources!() :: [{module(), map()}]

Returns the list of sources to load, in precedence order.

These are build-time inputs to LLMDB.Engine. The engine processes only the configured sources, in precedence order; it does not automatically merge the packaged runtime snapshot into the source list.

Configuration

config :llm_db,
  sources: [
    {LLMDB.Sources.ModelsDev, %{}},
    {LLMDB.Sources.Local, %{dir: "priv/llm_db"}}
  ]

Default Behavior

If not configured, returns an empty list [], meaning no build-time sources are configured. Runtime LLMDB.load/1 independently defaults to the packaged, version-pinned snapshot.

Returns

List of {module, opts} tuples in precedence order (first = lowest precedence).