BeamWeaver Google

BeamWeaver includes a Gemini Developer API provider under BeamWeaver.Google.

Implemented

• BeamWeaver.Google.ChatModel implements BeamWeaver.Core.ChatModel.

• Public model identifiers use the google: provider prefix, for example google:gemini-3.8-flash.

• Bare gemini-* identifiers are intentionally rejected so Gemini Developer API and future Vertex AI adapters do not share an ambiguous namespace.

• Requests go through BeamWeaver.Transport, so fake and replay transports can exercise provider behavior without live credentials.

• Namespace constructors load defaults from config :beam_weaver, :google; put any OS environment reads in your config/runtime.exs. Custom routing can use configured :base_url, explicit :base_url, or :endpoint.

• BeamWeaver messages become Gemini contents plus top-level systemInstruction.

• Custom tools become Gemini function declarations. Google built-ins such as Google Search, Google Maps, URL context, code execution, File Search, MCP servers, and model-specific computer use are pass-through provider request values.

• Gemini function parameter schemas are provider-sanitized: local $ref entries are dereferenced, unsupported JSON Schema annotation/object keywords such as $defs, title, default, and additionalProperties are removed, and nested property schemas are cleaned recursively.

• Checked-in profiles include google:gemini-3.8-flash, google:gemini-3.7-flash, google:gemini-3.6-flash, google:gemini-3.5-flash, google:gemini-3.5-flash-lite, google:gemini-3.1-flash-lite, and google:gemini-3.1-pro-preview. They accept text, image, video, audio, and PDF input and support thinking, structured output, function calling, caching, batch, flex, and priority inference.

• Built-in tools are profile-qualified. The 3.8, 3.7, 3.6, 3.5, and 3.5 Flash-Lite profiles include preview computer use. Gemini 3.1 Flash-Lite and 3.1 Pro Preview explicitly do not; callers get profile validation rather than a provider-side surprise.

• These models use thinking_level instead of thinking_budget. Gemini 3.8 Flash supports low, medium, and high and defaults to medium; Gemini 3.7 and 3.6 Flash also default to medium, while Gemini 3.5 Flash-Lite defaults to minimal. Google deprecates temperature, top_p, and top_k for these releases and does not support candidate_count, so their checked-in profiles reject those options before transport. They also reject requests whose last non-empty content turn has the Gemini model role; append a user or tool result turn instead of prefilling model output.

• Responses include normalized usage, reasoning/thinking token metadata, safety ratings, grounding metadata, model version, request IDs, and raw provider metadata.

• Streaming supports text deltas, typed stream envelopes, and reconstructed final assistant messages.

• Checked-in model profiles cover current recommended Gemini chat models. Stable Gemini 2.5 Pro, Flash, and Flash-Lite identifiers remain accepted through the explicit google:gemini-* family fallback. Retired preview variants, Gemini 2.0 Flash, and Gemini 3 Flash Preview return a :deprecated_model error with replacement and shutdown metadata.

• Gemini 3.5 Flash Cyber is not registered as a callable chat model. Google limits it to governments and trusted partners through CodeMender rather than exposing it through the Gemini Developer API, so its slug returns an :unsupported_model error before family fallback. See Google's release announcement for the stated rollout boundary.

Current Flash Profiles

Prices are USD per one million tokens for the Gemini Developer API. Cached input is the context-caching token price. Gemini 3.8, 3.7, and 3.6 Flash use Google's introductory pricing through December 31, 2026; their rates double on January 1, 2027. Their introductory cache storage price is $0.50 per one million tokens per hour.

For Gemini 3.1 Flash-Lite, audio input is priced separately at $0.50 standard, $0.25 batch/flex, and $0.90 priority per million audio tokens. Cached audio input is $0.05, $0.025, and $0.09 respectively. Standard cache storage is $1.00 per million tokens per hour and priority storage is $1.80. The profile records a scheduled shutdown on May 7, 2027 with Gemini 3.5 Flash-Lite as its recommended replacement.

From January 1, 2027, Gemini 3.8, 3.7, and 3.6 Flash standard input, cached input, and output rates become $1.50, $0.15, and $7.50 respectively. See the live pricing page for the corresponding batch, flex, priority, and storage rates.

See Google's model specifications and Gemini Developer API pricing for the live provider source of truth.

Usage

{:ok, model} =
  BeamWeaver.Models.init_chat_model("google:gemini-3.8-flash",
    thinking_level: :medium,
    include_thoughts: true
  )

BeamWeaver.Core.ChatModel.invoke(model, [
  BeamWeaver.Core.Message.user("Summarize the tradeoffs in one paragraph.")
])

Provider tools are request values:

tools = [
  BeamWeaver.Google.Tools.google_search(),
  BeamWeaver.Google.Tools.google_maps(),
  BeamWeaver.Google.Tools.code_execution(),
  BeamWeaver.Google.Tools.file_search(["fileSearchStores/my_store"]),
  my_local_tool
]

BeamWeaver.Core.ChatModel.invoke(model, messages, tools: tools, tool_choice: :auto)

Structured output maps to Gemini generation config:

BeamWeaver.Core.ChatModel.invoke(model, messages,
  response_format: %{
    schema: %{
      type: :object,
      properties: %{answer: %{type: :string}},
      required: [:answer]
    }
  }
)

When a structured-output request does not set :max_output_tokens, BeamWeaver uses the model profile's output limit for Gemini. For example, google:gemini-3.8-flash defaults structured-output calls to 65_536 maxOutputTokens, while an explicit max_output_tokens: value still wins.

Token counting uses Gemini's count-tokens endpoint:

BeamWeaver.Google.ChatModel.count_tokens(model, [
  BeamWeaver.Core.Message.user("Count this.")
])

Typed Streaming And Tools

Google typed streaming emits live text, reasoning, and custom tool-call events. After a successful stream, it emits exactly one %Events.Message{} followed by %Events.Done{}. The final message contains complete tool_calls, normalized usage_metadata, finish and provider metadata, and thought signatures needed for subsequent tool-result turns. It is decoded from the same streaming request, including any usage-only trailer.

alias BeamWeaver.Core.{ChatModel, Message, Tool}
alias BeamWeaver.Stream.{Envelope, Events}

ping =
  Tool.from_function!(
    name: "benchmark_ping",
    description: "Return pong.",
    input_schema: %{type: :object, properties: %{}},
    handler: fn _args, _opts -> "pong" end
  )

{:ok, events} =
  ChatModel.stream_typed_events(model, [Message.user("Call benchmark_ping.")],
    tools: [ping],
    tool_choice: :required
  )

Enum.each(events, fn
  %Envelope{event: %Events.Token{text: text}} ->
    IO.write(text)

  %Envelope{event: %Events.Message{message: message}} ->
    IO.inspect(message.tool_calls, label: "complete tool calls")
    IO.inspect(message.usage_metadata, label: "usage")

  %Events.Error{error: error} ->
    IO.inspect(error, label: "stream failed")

  _event ->
    :ok
end)

Standalone model calls return tool requests; agents execute them. Build an agent with model_opts: [stream: true] to consume the typed model stream and execute its completed calls. Preserve the final assistant message in history when sending tool results back to Google.

Google's custom :tool_call_delta events are live notifications, not MessageChunk.tool_call_chunks. Read executable calls from the final message. Treat that message as a complete snapshot; appending its text after displaying token deltas would duplicate the text.

Errors can arrive during enumeration, even after live deltas have been emitted. Provider, transport, malformed-SSE, and incomplete-stream errors emit %Events.Error{} and prevent a successful final Message or Done.

Unsupported Google Surfaces

• Vertex AI. That should be a separate explicit adapter/prefix rather than an alias of google:*.

• Dedicated image, audio, and video generation model modules beyond chat response modality options.

• Exact Python class identity and serialization compatibility. BeamWeaver keeps native Elixir modules, structs, and tagged errors.