Contributing

Please feel free to submit an feature request/issues/clarification items you ahve under Issues! _.


GenAI Core Library

genai_core is the foundation package of the Noizu GenAI family (alongside genai, genai_approval, and ex_llama): the protocols, threads, classifiers, pipelines, and tool infrastructure shared by Noizu Elixir AI applications.

The current architecture lives under lib/vnext_genai/ (the "VNext" branch-by-abstraction layer). The modules it exposes are GenAI.*, and the older lib/genai/ structures (graph, inference provider, model metadata, dialogue) remain as the legacy compatibility layer. New providers, settings, and capabilities merge in through extender protocols — thread, model, encoder, message/content, tool, and stream handler — without touching call sites, while the GenAI entry module stays a stable surface for chat, streaming, classification, media generation, and tool-chain execution.

Hosted providers (OpenAI, Anthropic, Gemini) are supported directly; local model support is provided by the ex_llama sibling package, which builds on genai_core.

Installation

def deps do
  [
    {:genai_core, "~> 0.4.0"}
  ]
end

Classifiers and conversation pipelines

GenAI.classify/2 supports typed choice, boolean and score decisions through explicit classifier adapters. GenAI.Pipeline adds extraction, inference-assisted preparation, transformations, result insertion, and isolated fork/join processing with a global concurrency budget across nested branches.

See the classifier pipeline guide for standalone and inline classification, agent postprocessing, evaluation forks and adapter contracts. The existing GenAI.run/1-3 and tool-chain behavior is preserved.

Forks are GenAI.Thread.Fork thread values: normal message/model/settings APIs, inference, streaming and tool chains dispatch to isolated leaves. Results carry ordered full branch paths, and partial failures retain successful siblings.

Threads: standard and session

GenAI.chat/2 opens a thread. :default and :standard create a GenAI.Thread.Standard — a linear conversation thread holding messages, settings, and model constraints in a compact state struct. :session creates a GenAI.Thread.Session, the advanced VNext thread: sessions track directives and a graph of nodes/links, carry runtime state, and surface optimization and grid-search loops through GenAI.execute/4 / GenAI.report/3.

Both thread types share the same builder and dispatch API — GenAI.with_model/2, GenAI.with_message/2-3, GenAI.with_setting/2-3, GenAI.run/1-3, and friends — routed through GenAI.ThreadProtocol, so code written against one thread type works against the other.

Streaming

GenAI.stream/3 runs inference in streaming mode; interstitial (dynamic) messages go to the thread's stream handler. Handlers implement GenAI.StreamHandler.Behaviour (begin stream, event handling, finish normalization) and can be overridden per thread with GenAI.with_stream_handler/3. Built-in handlers decode provider SSE with a pure incremental feed/2 step — Anthropic and OpenAI normalization is built in (GenAI.StreamHandler.SSE accumulates the final GenAI.ChatCompletion as events arrive).

Prompt cache usage

Since 0.3.5, GenAI.ChatCompletion.Usage carries cache_read_input_tokens / cache_creation_input_tokens. OpenAI-compatible responses map usage.prompt_tokens_details.cached_tokens to the read field, and the Anthropic and OpenAI stream handlers pass cache usage through to accumulated completions. Caching is requested per content part via the optional cache_control field on GenAI.Message.Content.TextContent (consumed by the genai Anthropic encoder).

Tool sources and bounded chains

GenAI.Tool.Registry lets an application expose tools from MCP clients, in-process harnesses, databases, or other sources without adding those dependencies to genai_core. An adapter implements two callbacks:

defmodule MyApp.MCPTools do
  @behaviour GenAI.Tool.Source

  def list_tools(client, context, options) do
    # Return GenAI.Tool values or MCP-shaped maps with name, description, and
    # inputSchema/parameters.
    MyMCPClient.list_tools(client, context, options)
  end

  def call_tool(client, name, arguments, context, options) do
    MyMCPClient.call_tool(client, name, arguments, context, options)
  end
end

{:ok, registry} =
  GenAI.Tool.Registry.new()
  |> GenAI.Tool.Registry.register(:project_mcp, MyApp.MCPTools, client, context)

{:ok, {completion, thread, summary}} =
  GenAI.run_with_tools(thread, registry, context,
    max_iterations: 8,
    max_tool_calls: 32,
    timeout: 15_000
  )

Model-facing names are deterministic and namespaced by default, for example project_mcp__search. Unsupported characters become _, and a normalized-name collision returns {:error, {:tool_name_collision, name}}. Callers must opt in to collision: :keep or collision: :replace to choose another policy.

The chain runner is explicit and does not change GenAI.run/1-3. It caps configured limits at 64 inference iterations, 256 total tool calls, and 120 seconds per call. Source exceptions, exits, timeouts, calls, and results are normalized through GenAI.Tool.Call and GenAI.Tool.Result.

Telemetry and payload safety

Lifecycle events are emitted under these prefixes when the optional :telemetry module is available:

Consumers without :telemetry can pass a three-argument function using the telemetry: option. Metadata contains names, IDs, status, bounds, and counts; tool arguments and results are excluded by default. Set include_payloads: true only where those values are safe to observe.

Media generation (ADR-016)

GenAI.generate_media/2 routes a GenAI.Media.Request to a provider that declares support for its (input, output) modality and runs its generate_media/2. Modalities cover text -> image/speech/music/sfx/video and speech -> text (transcription). Sync providers return {:ok, %{data: binary, mime: String.t(), meta: map}}; long-running providers (music/video) return {:ok, %GenAI.Media.Job{}} for submit -> poll -> download.

Providers register through an explicit modality registry — config :genai, :media_providers, listing modules that implement supported_modalities/0 — so GenAI.Media.Router never hardcodes capabilities. Free-form chat/vision (image input -> text generation) is not this path; it rides the normal chat/thread run via the encoder protocol.

Graphs and Mermaid

Session threads model advanced flows as graphs of nodes and links (GenAI.Graph.Node, GenAI.Graph.Link) with a matching protocol for rendering a graph as a Mermaid diagram (GenAI.Graph.MermaidProtocol), useful for inspecting execution plans of optimization and grid-search loops.