217 lines
6.5 KiB
Rust
217 lines
6.5 KiB
Rust
use crate::mcp::llmclient::{LLMClient, LLMError};
|
|
use async_trait::async_trait;
|
|
use ollama_rs::generation::chat::ChatMessage as OllamaChatMessage;
|
|
use ollama_rs::generation::chat::request::ChatMessageRequest as OllamaChatMessageRequest;
|
|
use ollama_rs::generation::chat::{ChatMessageResponse, MessageRole};
|
|
use ollama_rs::generation::tools::{ToolCall, ToolInfo};
|
|
use ollama_rs::history::ChatHistory;
|
|
use serde::{Deserialize, Serialize};
|
|
|
|
#[derive(Debug, Clone)]
|
|
pub struct Mistral {
|
|
api_key: String,
|
|
system_prompt: Option<String>,
|
|
}
|
|
|
|
impl Mistral {
|
|
pub fn new(api_key: String) -> Mistral {
|
|
Mistral {
|
|
api_key,
|
|
system_prompt: None,
|
|
}
|
|
}
|
|
|
|
pub fn with_system_prompt(mut self, system_prompt: Option<String>) -> Mistral {
|
|
self.system_prompt = system_prompt;
|
|
self
|
|
}
|
|
}
|
|
|
|
#[derive(Debug, Clone, Deserialize, Serialize)]
|
|
pub struct Message {
|
|
content: String,
|
|
role: MessageRole,
|
|
#[serde(
|
|
skip_serializing_if = "Vec::is_empty",
|
|
default = "Vec::new",
|
|
deserialize_with = "parse_tool_calls"
|
|
)]
|
|
tool_calls: Vec<ToolCall>,
|
|
}
|
|
|
|
fn parse_tool_calls<'de, D>(d: D) -> Result<Vec<ToolCall>, D::Error>
|
|
where
|
|
D: serde::Deserializer<'de>,
|
|
{
|
|
Deserialize::deserialize(d).map(|x: Option<_>| x.unwrap_or(Vec::new()))
|
|
}
|
|
|
|
impl From<OllamaChatMessage> for Message {
|
|
fn from(message: OllamaChatMessage) -> Message {
|
|
Message {
|
|
content: message.content,
|
|
role: message.role,
|
|
tool_calls: message.tool_calls,
|
|
}
|
|
}
|
|
}
|
|
|
|
impl From<Message> for OllamaChatMessage {
|
|
fn from(message: Message) -> OllamaChatMessage {
|
|
OllamaChatMessage {
|
|
role: message.role,
|
|
content: message.content,
|
|
tool_calls: message.tool_calls,
|
|
images: None,
|
|
thinking: None,
|
|
}
|
|
}
|
|
}
|
|
|
|
#[derive(Debug, Serialize)]
|
|
struct ChatCompletionRequest {
|
|
messages: Vec<Message>,
|
|
model: String,
|
|
tools: Vec<ToolInfo>,
|
|
}
|
|
|
|
impl From<OllamaChatMessageRequest> for ChatCompletionRequest {
|
|
fn from(request: OllamaChatMessageRequest) -> ChatCompletionRequest {
|
|
ChatCompletionRequest {
|
|
messages: request.messages.into_iter().map(|m| m.into()).collect(),
|
|
model: request.model_name,
|
|
tools: request.tools,
|
|
}
|
|
}
|
|
}
|
|
|
|
#[derive(Debug, Clone, Deserialize)]
|
|
struct ChatCompletionChoice {
|
|
message: Message,
|
|
}
|
|
|
|
#[derive(Debug, Clone, Deserialize)]
|
|
struct ChatCompletionResponse {
|
|
choices: Vec<ChatCompletionChoice>,
|
|
model: String,
|
|
#[serde(rename = "created")]
|
|
created_unix_seconds: i64,
|
|
}
|
|
|
|
impl TryFrom<ChatCompletionResponse> for ChatMessageResponse {
|
|
type Error = LLMError;
|
|
|
|
fn try_from(value: ChatCompletionResponse) -> Result<ChatMessageResponse, LLMError> {
|
|
let created = chrono::DateTime::from_timestamp(value.created_unix_seconds, 0)
|
|
.ok_or(LLMError::Other("failed converting timestamp".to_string()))?;
|
|
|
|
Ok(ChatMessageResponse {
|
|
model: value.model,
|
|
created_at: created.to_rfc3339(),
|
|
message: value
|
|
.choices
|
|
.first()
|
|
.ok_or(LLMError::Other("answer does not exist".to_string()))?
|
|
.message
|
|
.clone()
|
|
.into(),
|
|
logprobs: None,
|
|
done: false,
|
|
final_data: None,
|
|
})
|
|
}
|
|
}
|
|
|
|
#[async_trait]
|
|
impl LLMClient for Mistral {
|
|
async fn send_chat_messages_with_history(
|
|
&self,
|
|
history: &mut Vec<OllamaChatMessage>,
|
|
mut request: ollama_rs::generation::chat::request::ChatMessageRequest,
|
|
) -> Result<ChatMessageResponse, LLMError> {
|
|
if let Some(first_message) = history.first() // add system prompt as first message
|
|
&& first_message.role == MessageRole::System
|
|
&& let Some(system_prompt) = self.system_prompt.clone()
|
|
{
|
|
history.insert(
|
|
0,
|
|
OllamaChatMessage::new(MessageRole::System, system_prompt),
|
|
);
|
|
}
|
|
|
|
history.append(&mut request.messages);
|
|
|
|
request.messages = history.messages().to_vec();
|
|
|
|
let client = reqwest::Client::new();
|
|
let chat_completion_request: ChatCompletionRequest = request.into();
|
|
let response = client
|
|
.post("https://api.mistral.ai/v1/chat/completions")
|
|
.header("Authorization", format!("Bearer {}", self.api_key))
|
|
.json(&chat_completion_request)
|
|
.send()
|
|
.await
|
|
.map_err(|e| LLMError::NetworkError(e.to_string()))?;
|
|
|
|
let response = response
|
|
.error_for_status()
|
|
.map_err(|e| LLMError::Other(e.to_string()))?;
|
|
|
|
let completion_response: ChatCompletionResponse = response
|
|
.json()
|
|
.await
|
|
.map_err(|e| LLMError::Other(e.to_string()))?;
|
|
let chat_message_response: ChatMessageResponse = completion_response.try_into()?;
|
|
|
|
Ok(chat_message_response)
|
|
}
|
|
|
|
async fn preload_model(&self, _model: &str) {
|
|
// do nothing
|
|
}
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod tests {
|
|
use super::*;
|
|
use ollama_rs::generation::chat::request::ChatMessageRequest;
|
|
use ollama_rs::generation::chat::{ChatMessage, MessageRole};
|
|
use rmcp::schemars;
|
|
|
|
#[derive(schemars::JsonSchema)]
|
|
struct BeepToolParameters {
|
|
location: String,
|
|
}
|
|
|
|
#[tokio::test]
|
|
async fn sending_message() {
|
|
let client = Mistral::new(include_str!(".MISTRAL_API_KEY").to_string());
|
|
|
|
let mut history = vec![];
|
|
|
|
dbg!(
|
|
client
|
|
.send_chat_messages_with_history(
|
|
&mut history,
|
|
ChatMessageRequest::new(
|
|
"mistral-medium-3-5".to_string(),
|
|
vec![ChatMessage::new(
|
|
MessageRole::User,
|
|
"Mach ein Geräusch in der Küche".to_string()
|
|
)],
|
|
)
|
|
.tools(vec![ToolInfo {
|
|
tool_type: ollama_rs::generation::tools::ToolType::Function,
|
|
function: ollama_rs::generation::tools::ToolFunctionInfo {
|
|
name: "make_sound".to_string(),
|
|
description: "alerts the user with a beep".to_string(),
|
|
parameters: schemars::schema_for!(BeepToolParameters),
|
|
}
|
|
}]),
|
|
)
|
|
.await
|
|
.unwrap()
|
|
);
|
|
}
|
|
}
|