Mnemosyne.Config (mnemosyne v0.1.6)

Unified configuration for Mnemosyne LLM and embedding settings.

Holds default model configuration for LLM and embedding calls, plus per-step overrides that merge on top of defaults. Configuration is validated at load time using Zoi schemas to catch misconfiguration early.

Fields

:session (map/0) - Required. Session auto-commit and timeout configuration The default value is %{auto_commit: true, flush_timeout_ms: 120000, session_timeout_ms: 600000}.
:overrides (map/0) - Required. Per-step LLM overrides keyed by pipeline step atom The default value is %{}.
:embedding (map/0) - Required. Default embedding model configuration for vector generation
:backend (map/0) - Graph backend configuration
:llm (map/0) - Required. Default LLM configuration applied to all pipeline steps
:value_function (map/0) - Required. Value function module and per-node-type scoring parameters The default value is %{module: Mnemosyne.ValueFunction.Default, params: %{tag: %{k: 5, threshold: 0.9, lambda: 0.01, top_k: 10, base_floor: 0.3, beta: 1.0}, source: %{k: 5, threshold: 0.0, lambda: 0.01, top_k: 50, base_floor: 0.3, beta: 1.0}, semantic: %{k: 5, threshold: 0.0, lambda: 0.01, top_k: 20, base_floor: 0.3, beta: 1.0}, procedural: %{k: 5, threshold: 0.8, lambda: 0.01, top_k: 10, base_floor: 0.3, beta: 1.0}, episodic: %{k: 5, threshold: 0.0, lambda: 0.01, top_k: 30, base_floor: 0.3, beta: 1.0}, subgoal: %{k: 5, threshold: 0.75, lambda: 0.01, top_k: 10, base_floor: 0.3, beta: 1.0}, intent: %{k: 5, threshold: 0.7, lambda: 0.01, top_k: 10, base_floor: 0.3, beta: 1.0}}}.
:episodic_validation (map/0) - Required. The default value is nil.
:intent_merge_threshold (float/0) - Required. The default value is 0.8.
:intent_identity_threshold (float/0) - Required. The default value is 0.95.
:refinement_threshold (float/0) - Required. The default value is 0.6.
:refinement_budget (integer/0) - Required. The default value is 1.
:plateau_delta (float/0) - Required. The default value is 0.05.
:trace_verbosity (:summary | :detailed) - Required. The default value is :summary.
:extraction_profile (term/0) - Required. The default value is nil.

Override Resolution

When a pipeline step (e.g. :structuring, :retrieval) has an entry in :overrides, the override's :model replaces the default and its :opts are deep-merged with the base LLM opts. This lets you use a cheaper model for simple extraction steps while keeping a powerful model for reasoning.

Examples

Minimal configuration with defaults:

config = %Mnemosyne.Config{
  llm: %{model: "gpt-4o", opts: %{}},
  embedding: %{model: "text-embedding-3-small", opts: %{}}
}

With per-step overrides:

config = %Mnemosyne.Config{
  llm: %{model: "gpt-4o", opts: %{temperature: 0.7}},
  embedding: %{model: "text-embedding-3-small", opts: %{}},
  overrides: %{
    structuring: %{model: "gpt-4o-mini"},
    retrieval: %{opts: %{temperature: 0.0}}
  }
}

Loading from application environment:

# In config/config.exs
config :mnemosyne, :config,
  llm: %{model: "gpt-4o", opts: %{temperature: 0.7}},
  embedding: %{model: "text-embedding-3-small", opts: %{}}

# At runtime
{:ok, config} = Mnemosyne.Config.from_env()

Summary

Types

t()

Functions

embedding_opts(config)

Returns embedding keyword opts from the config as a flat keyword list.

from_env()

Loads and validates config from the :mnemosyne application environment.

llm_opts(config, step, base_opts)

Returns LLM keyword opts for a pipeline step, merging overrides with base opts.

resolve(config, step)

Returns the LLM model and opts for the given pipeline step, applying any overrides.

resolve_embedding(config)

Returns the embedding model and opts from the config.

resolve_overlay(arg1, step)

Returns the overlay text for a given pipeline step, or nil if no profile or overlay exists.

resolve_value_function(config, node_type)

Returns the value function params for a given node type.

t()

Types

t()

@type t() :: %Mnemosyne.Config{
  backend: nil | %{module: atom(), opts: map()},
  embedding: %{opts: map(), model: binary()},
  episodic_validation: nil | map(),
  extraction_profile: nil | any(),
  intent_identity_threshold: float(),
  intent_merge_threshold: float(),
  llm: %{opts: map(), model: binary()},
  overrides: %{
    optional(atom()) => %{
      optional(:opts) => map(),
      optional(:model) => binary()
    }
  },
  plateau_delta: float(),
  refinement_budget: integer(),
  refinement_threshold: float(),
  session: %{
    auto_commit: boolean(),
    flush_timeout_ms: integer() | :infinity,
    session_timeout_ms: integer() | :infinity
  },
  trace_verbosity: :summary | :detailed,
  value_function: %{
    module: atom(),
    params: %{
      optional(atom()) => %{
        k: integer(),
        threshold: float(),
        lambda: float(),
        top_k: integer(),
        base_floor: float(),
        beta: float()
      }
    }
  }
}

Functions

embedding_opts(config)

@spec embedding_opts(t() | nil) :: keyword()

Returns embedding keyword opts from the config as a flat keyword list.

Returns an empty list when config is nil, allowing callers to work without configuration.

from_env()

@spec from_env() :: {:ok, t()} | {:error, Mnemosyne.Errors.Invalid.ConfigError.t()}

Loads and validates config from the :mnemosyne application environment.

Reads the :config key under the :mnemosyne application and validates it against the Zoi schema. Returns {:error, ConfigError} when the key is missing or the data fails validation.

Examples

# When configured:
{:ok, %Mnemosyne.Config{}} = Mnemosyne.Config.from_env()

# When missing:
{:error, %Mnemosyne.Errors.Invalid.ConfigError{reason: :no_config}} = Mnemosyne.Config.from_env()

llm_opts(config, step, base_opts)

@spec llm_opts(t() | nil, atom(), keyword()) :: keyword()

Returns LLM keyword opts for a pipeline step, merging overrides with base opts.

Resolves the model and opts for the given step, then flattens the result into a keyword list suitable for passing directly to an LLM adapter call. The base_opts (typically pipeline-specific options like prompt messages) are appended after the resolved config options.

Returns base_opts unchanged when config is nil, allowing callers to work without configuration.

resolve(config, step)

@spec resolve(t(), atom()) :: %{model: String.t(), opts: map()}

Returns the LLM model and opts for the given pipeline step, applying any overrides.

Looks up the step atom in config.overrides. If an override exists, its :model replaces the base model (when present) and its :opts are merged on top of the base opts. When no override exists, the base LLM config is returned as-is.

Examples

iex> config = %Mnemosyne.Config{
...>   llm: %{model: "gpt-4o", opts: %{temperature: 0.7}},
...>   embedding: %{model: "e5-base-v2", opts: %{}},
...>   overrides: %{structuring: %{model: "gpt-4o-mini"}}
...> }
iex> Mnemosyne.Config.resolve(config, :structuring)
%{model: "gpt-4o-mini", opts: %{temperature: 0.7}}
iex> Mnemosyne.Config.resolve(config, :retrieval)
%{model: "gpt-4o", opts: %{temperature: 0.7}}

resolve_embedding(config)

@spec resolve_embedding(t()) :: %{model: String.t(), opts: map()}

Returns the embedding model and opts from the config.

Unlike resolve/2, embeddings have no per-step overrides since the same embedding model must be used consistently across the entire knowledge graph to keep vector spaces comparable.

resolve_overlay(arg1, step)

@spec resolve_overlay(t() | nil, atom()) :: String.t() | nil

Returns the overlay text for a given pipeline step, or nil if no profile or overlay exists.

resolve_value_function(config, node_type)

@spec resolve_value_function(t(), atom()) :: map()

Returns the value function params for a given node type.

Looks up the type in config.value_function.params and merges with per-type defaults. Returns safe defaults for unknown types.