new: implemented google gemini integration (closes #33)

docs/index.md

@@ -75,11 +75,14 @@ Nerve features integrations for any model accessible via the following providers
 | **xAI** | `XAI_API_KEY` | `xai://grok-beta` |
 | **Mistral.ai** | `MISTRAL_API_KEY` | `mistral://mistral-large-latest` |
 | **Novita** | `NOVITA_API_KEY` | `novita://meta-llama/llama-3.1-70b-instruct` |
+| **Google Gemini**³ | `GEMINI_API_KEY` | `gemini://gemini-2.0-flash` |
 
 ¹ **o1-preview and o1 models do not support function calling directly** and do not support a system prompt. Nerve will try to detect this and fallback to user prompt. It is possible to force this behaviour by adding the `--user-only` flag to the command line.
 
 ² Refer to [this document](https://huggingface.co/blog/tgi-messages-api#using-inference-endpoints-with-openai-client-libraries) for how to configure a custom Huggingface endpoint.
 
+³ Google Gemini OpenAI endpoint [breaks with multiple tools](https://discuss.ai.google.dev/t/invalid-argument-error-using-openai-compatible/51788). While this bug won't be fixed, Nerve will detect this and use its own xml based tooling prompt to work around this issue.
+
 ### Using with Robopages
 
 Nerve can use functions from a [robopages server](https://github.com/dreadnode/robopages-cli). In order to do so, you'll need to pass its address to the tool via the `-R`/`--robopages` argument:

src/agent/generator/google.rs

@@ -0,0 +1,45 @@
+use anyhow::Result;
+use async_trait::async_trait;
+
+use crate::agent::state::SharedState;
+
+use super::{openai::OpenAIClient, ChatOptions, ChatResponse, Client, SupportedFeatures};
+
+pub struct GoogleClient {
+    client: OpenAIClient,
+}
+
+#[async_trait]
+impl Client for GoogleClient {
+    fn new(_: &str, _: u16, model_name: &str, _: u32) -> anyhow::Result<Self>
+    where
+        Self: Sized,
+    {
+        let client = OpenAIClient::custom(
+            model_name,
+            "GEMINI_API_KEY",
+            "https://generativelanguage.googleapis.com/v1beta/openai/",
+        )?;
+
+        Ok(Self { client })
+    }
+
+    async fn check_supported_features(&self) -> Result<SupportedFeatures> {
+        self.client.check_supported_features().await
+    }
+
+    async fn chat(
+        &self,
+        state: SharedState,
+        options: &ChatOptions,
+    ) -> anyhow::Result<ChatResponse> {
+        self.client.chat(state, options).await
+    }
+}
+
+#[async_trait]
+impl mini_rag::Embedder for GoogleClient {
+    async fn embed(&self, text: &str) -> Result<mini_rag::Embeddings> {
+        self.client.embed(text).await
+    }
+}

src/agent/generator/mod.rs

@@ -13,6 +13,7 @@ use super::{namespaces::ActionOutput, state::SharedState, Invocation};
 mod anthropic;
 mod deepseek;
 mod fireworks;
+mod google;
 mod groq;
 mod huggingface;
 mod mistral;
@@ -238,6 +239,12 @@ macro_rules! factory_body {
                 $model_name,
                 $context_window,
             )?)),
+            "google" | "gemini" => Ok(Box::new(google::GoogleClient::new(
+                $url,
+                $port,
+                $model_name,
+                $context_window,
+            )?)),
             "http" => Ok(Box::new(openai_compatible::OpenAiCompatibleClient::new(
                 $url,
                 $port,

src/agent/generator/openai.rs

@@ -27,6 +27,7 @@ pub struct OpenAiToolFunctionParameters {
 }
 
 pub struct OpenAIClient {
+    ident: String,
     model: String,
     client: OpenAI,
 }
@@ -38,21 +39,35 @@ impl OpenAIClient {
     {
         let model = model.to_string();
         let api_key = std::env::var(api_key_env).map_err(|_| anyhow!("Missing {api_key_env}"))?;
+        let ident = api_key_env
+            .split('_')
+            .next()
+            .unwrap_or("openai")
+            .to_string();
         let auth = Auth::new(&api_key);
         let client = OpenAI::new(auth, endpoint);
 
-        Ok(Self { model, client })
+        Ok(Self {
+            ident,
+            model,
+            client,
+        })
     }
 
     pub fn custom_no_auth(model: &str, endpoint: &str) -> anyhow::Result<Self>
     where
         Self: Sized,
     {
+        let ident = "http".to_string();
         let model = model.to_string();
         let auth = Auth::new("");
         let client = OpenAI::new(auth, endpoint);
 
-        Ok(Self { model, client })
+        Ok(Self {
+            ident,
+            model,
+            client,
+        })
     }
 
     async fn get_tools_if_supported(&self, state: &SharedState) -> Vec<FunctionTool> {
@@ -187,6 +202,14 @@ impl Client for OpenAIClient {
                 && api_error.contains("does not support 'system' with this model")
             {
                 system_prompt_support = false;
+            } else if self.ident == "GEMINI" && api_error.contains("INVALID_ARGUMENT") {
+                // Gemini openai endpoint breaks with multiple tools:
+                //
+                //  https://discuss.ai.google.dev/t/invalid-argument-error-using-openai-compatible/51788
+                //  https://discuss.ai.google.dev/t/gemini-openai-compatibility-multiple-functions-support-in-function-calling-error-400/49431
+                //
+                // Never can overcome this bug by providing its own xml based tooling prompt.
+                log::warn!("this is a documented bug of Google Gemini OpenAI endpoint: https://discuss.ai.google.dev/t/invalid-argument-error-using-openai-compatible/51788");
             } else {
                 log::error!("openai.check_tools_support.error = {}", api_error);
             }