🔧 Allow backend agnostic configuration (#23)

Author: chrisliebaer

Cargo.toml

@@ -27,7 +27,7 @@ rand = "0.9"
 humantime = "2.1"
 log = "0.4"
 
-llm = { version = "1.3", features = ["openai", "rustls-tls"] }
+llm = { version = "1.3", features = ["rustls-tls"] }
 rmcp = { version = "0.6.0", features = [
     "client",
     "transport-sse-client-reqwest",

README.md

@@ -22,14 +22,16 @@ The bot is in a very early stage of development but already usable.
 
 The project requires the following environment variables:
 
-- `OPENAI_TOKEN`: Your OpenAI API token.
+- `API_KEY`: Your LLM provider API key.
+- `LLM_PROVIDER`: The LLM provider to use. Supported values: `openai`, `anthropic`, `ollama`, `deepseek`, `xai`, `phind`, `google`, `groq`, `azureopenai`, `elevenlabs`, `cohere`, `mistral`, `openrouter`.
 - `MODEL`: The model to use.
 - `DISCORD_TOKEN`: Your Discord bot token.
 - `TEMPLATE_DIR`: The directory where your Tera templates are located. Defaults to `templates`.
 - `RATE_LIMIT_CONFIG`: The path to your rate limit configuration file. Defaults to `rate_limits.toml`.
 - `DATABASE_URL`: The URL to your database. For example `mysql://user:password@localhost/database`.
 - `WHITELIST`: A comma separated list of Discord snowflakes for channels, categories, or guilds in which the bot should respond. If empty, the bot will respond in all channels. Defaults to an empty string.
 - `OPT_OUT_LOCKOUT`: The time in seconds a user is locked out from the bot after opting out. Defaults to `30d`. Can use any time format supported by the `humantime` crate.
+- `COMPLETION_TIMEOUT`: The timeout for LLM completion requests. Defaults to `60s`. Can use any time format supported by the `humantime` crate.
 
 ## License
 

src/handler/completion.rs

@@ -150,10 +150,7 @@ pub async fn handle_completion(
 
 	let typing_notification = typing_indicator(ctx, new_message.channel_id);
 
-	let completion_request = tokio::time::timeout(
-		std::time::Duration::from_secs(60),
-		generate_llm_response(ctx, app, new_message),
-	);
+	let completion_request = tokio::time::timeout(app.completion_timeout, generate_llm_response(ctx, app, new_message));
 
 	// assuming typing notifications don't fail, we can just wait for the fork to finish and will keep sending typing
 	// notifications in the meantime
@@ -366,10 +363,10 @@ async fn generate_llm_response<'a>(
 			}
 
 			// add assistant's tool call to conversation
-			conversation.push(ChatMessage::assistant().tool_use(tool_calls.clone()).content("").build());
+			conversation.push(ChatMessage::assistant().tool_use(tool_calls.clone()).build());
 
 			// add tool results to conversation
-			conversation.push(ChatMessage::assistant().tool_result(tool_results.clone()).content("").build());
+			conversation.push(ChatMessage::user().tool_result(tool_results.clone()).build());
 		} else {
 			// No tool calls - we have our final response
 			let content = response.text().ok_or(miette!("LLM response has no content"))?;

src/main.rs

@@ -103,8 +103,11 @@ lazy_static! {
 
 #[derive(Envconfig)]
 struct EnvConfig {
-	#[envconfig(from = "OPENAI_TOKEN")]
-	openai_token: String,
+	#[envconfig(from = "API_KEY")]
+	api_key: String,
+
+	#[envconfig(from = "LLM_PROVIDER")]
+	llm_provider: String,
 
 	#[envconfig(from = "MODEL")]
 	model: String,
@@ -126,6 +129,19 @@ struct EnvConfig {
 
 	#[envconfig(from = "WHITELIST", default = "")]
 	whitelist: Whitelist,
+
+	#[envconfig(from = "COMPLETION_TIMEOUT", default = "60s")]
+	completion_timeout: ParsedDuration,
+}
+
+impl EnvConfig {
+	/// Converts the provider string to LLMBackend enum
+	fn get_llm_backend(&self) -> Result<LLMBackend> {
+		use std::str::FromStr;
+		LLMBackend::from_str(&self.llm_provider)
+			.into_diagnostic()
+			.wrap_err_with(|| format!("unsupported LLM provider: {}", self.llm_provider))
+	}
 }
 
 struct ParsedDuration(Duration);
@@ -224,6 +240,7 @@ struct AppState {
 	context_settings: InvocationContextSettings,
 	whitelist: Whitelist,
 	opt_out_lockout: Duration,
+	completion_timeout: Duration,
 }
 
 type Context<'a> = poise::Context<'a, AppState, Report>;
@@ -248,9 +265,11 @@ async fn main() -> Result<()> {
 	};
 
 	let llm_client = {
+		let backend = env_config.get_llm_backend()?;
+		
 		let mut builder = LLMBuilder::new()
-			.backend(LLMBackend::OpenAI)
-			.api_key(&env_config.openai_token)
+			.backend(backend)
+			.api_key(&env_config.api_key)
 			.model(&env_config.model)
 			.max_tokens(2000);
 
@@ -355,6 +374,7 @@ async fn main() -> Result<()> {
 					},
 					whitelist: env_config.whitelist,
 					opt_out_lockout: env_config.opt_out_lockout.0,
+					completion_timeout: env_config.completion_timeout.0,
 				})
 			})
 		})

src/mcp_config.rs

@@ -112,12 +112,6 @@ mod tests {
 			},
 			_ => panic!("Expected HTTP server config"),
 		}
-
-		assert!(server.is_http_based());
-		assert_eq!(
-			server.get_connection_url(),
-			Some("http://192.168.200.10:8096/servers/web-search/sse")
-		);
 	}
 
 	/// Test parsing an SSE server configuration
@@ -153,8 +147,6 @@ mod tests {
 			},
 			_ => panic!("Expected SSE server config"),
 		}
-
-		assert!(server.is_http_based());
 	}
 
 	/// Test parsing a stdio server configuration
@@ -194,9 +186,6 @@ mod tests {
 			},
 			_ => panic!("Expected Stdio server config"),
 		}
-
-		assert!(!server.is_http_based());
-		assert_eq!(server.get_connection_url(), None);
 	}
 
 	/// Test parsing multiple servers with mixed types
@@ -228,12 +217,6 @@ mod tests {
 		assert!(config.servers.contains_key("web-search"));
 		assert!(config.servers.contains_key("web-fetch"));
 		assert!(config.servers.contains_key("local-tool"));
-
-		let http_count = config.servers.values().filter(|s| s.is_http_based()).count();
-		let stdio_count = config.servers.values().filter(|s| !s.is_http_based()).count();
-
-		assert_eq!(http_count, 2);
-		assert_eq!(stdio_count, 1);
 	}
 
 	/// Test parsing config similar to the provided mcp.json