From 73c516ab28f2025c6ab61fd5983349ba02fbac79 Mon Sep 17 00:00:00 2001 From: syui Date: Thu, 12 Jun 2025 21:42:30 +0900 Subject: [PATCH] fix openai tools --- .claude/settings.local.json | 3 +- .gitignore | 3 + src/openai_provider.rs | 247 +++++++++++++++++++++++++++++++----- src/persona.rs | 8 +- 4 files changed, 230 insertions(+), 31 deletions(-) diff --git a/.claude/settings.local.json b/.claude/settings.local.json index 65d81cf..36c0a34 100644 --- a/.claude/settings.local.json +++ b/.claude/settings.local.json @@ -52,7 +52,8 @@ "Bash(cargo check:*)", "Bash(cargo run:*)", "Bash(cargo test:*)", - "Bash(diff:*)" + "Bash(diff:*)", + "Bash(cargo:*)" ], "deny": [] } diff --git a/.gitignore b/.gitignore index 6536fb8..944b959 100644 --- a/.gitignore +++ b/.gitignore @@ -6,3 +6,6 @@ mcp/scripts/__* data __pycache__ conversations.json +json/*.zip +json/*/* +*.log diff --git a/src/openai_provider.rs b/src/openai_provider.rs index 7941ca2..0ff0b1d 100644 --- a/src/openai_provider.rs +++ b/src/openai_provider.rs @@ -35,9 +35,17 @@ impl OpenAIProvider { } } - pub fn with_system_prompt(mut self, prompt: String) -> Self { - self.system_prompt = Some(prompt); - self + pub fn with_system_prompt(api_key: String, model: Option, system_prompt: Option) -> Self { + let config = async_openai::config::OpenAIConfig::new() + .with_api_key(api_key); + let client = Client::with_config(config); + + Self { + client, + model: model.unwrap_or_else(|| "gpt-4".to_string()), + service_client: ServiceClient::new(), + system_prompt, + } } /// Generate OpenAI tools from MCP endpoints (matching Python implementation) @@ -48,7 +56,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "get_memories".to_string(), - description: Some("過去の会話記憶を取得します。「覚えている」「前回」「以前」などの質問で必ず使用してください".to_string()), + description: Some("Get past conversation memories".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -65,7 +73,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "search_memories".to_string(), - description: Some("特定のトピックについて話した記憶を検索します。「プログラミングについて」「○○について話した」などの質問で使用してください".to_string()), + description: Some("Search memories for specific topics or keywords".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -83,7 +91,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "get_contextual_memories".to_string(), - description: Some("クエリに関連する文脈的記憶を取得します".to_string()), + description: Some("Get contextual memories related to a query".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -105,7 +113,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "get_relationship".to_string(), - description: Some("特定ユーザーとの関係性情報を取得します".to_string()), + description: Some("Get relationship information with a specific user".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -123,7 +131,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "card_get_user_cards".to_string(), - description: Some("ユーザーが所有するカードの一覧を取得します".to_string()), + description: Some("Get user's card collection".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -145,7 +153,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "card_draw_card".to_string(), - description: Some("ガチャを引いてカードを取得します".to_string()), + description: Some("Draw a card from the gacha system".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -167,7 +175,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "card_analyze_collection".to_string(), - description: Some("ユーザーのカードコレクションを分析します".to_string()), + description: Some("Analyze user's card collection".to_string()), parameters: Some(json!({ "type": "object", "properties": { @@ -184,7 +192,7 @@ impl OpenAIProvider { r#type: ChatCompletionToolType::Function, function: FunctionObject { name: "card_get_gacha_stats".to_string(), - description: Some("ガチャの統計情報を取得します".to_string()), + description: Some("Get gacha statistics".to_string()), parameters: Some(json!({ "type": "object", "properties": {} @@ -200,9 +208,11 @@ impl OpenAIProvider { pub async fn chat_with_mcp(&self, prompt: String, user_id: String) -> Result { let tools = self.get_mcp_tools(); + let system_content = self.system_prompt.as_deref().unwrap_or( - "あなたは記憶システムと関係性データ、カードゲームシステムにアクセスできるAIです。\n\n【重要】以下の場合は必ずツールを使用してください：\n\n1. カード関連の質問:\n- 「カード」「コレクション」「ガチャ」「見せて」「持っている」「状況」「どんなカード」などのキーワードがある場合\n- card_get_user_cardsツールを使用してユーザーのカード情報を取得\n\n2. 記憶・関係性の質問:\n- 「覚えている」「前回」「以前」「について話した」「関係」などのキーワードがある場合\n- 適切なメモリツールを使用\n\n3. パラメータの設定:\n- didパラメータには現在会話しているユーザーのID（例：'syui'）を使用\n- ツールを積極的に使用して正確な情報を提供してください\n\nユーザーが何かを尋ねた時は、まず関連するツールがあるかを考え、適切なツールを使用してから回答してください。" + "You are an AI assistant with access to memory, relationship data, and card game systems. Use the available tools when appropriate to provide accurate and contextual responses." ); + let request = CreateChatCompletionRequestArgs::default() .model(&self.model) @@ -220,19 +230,21 @@ impl OpenAIProvider { } ), ]) - .tools(tools) + .tools(tools.clone()) .tool_choice(ChatCompletionToolChoiceOption::Auto) .max_tokens(2000u16) .temperature(0.7) .build()?; + let response = self.client.chat().create(request).await?; let message = &response.choices[0].message; + // Handle tool calls if let Some(tool_calls) = &message.tool_calls { if tool_calls.is_empty() { - println!("🔧 [OpenAI] No tools called"); + println!("🔧 [OpenAI] No tools called (empty array)"); } else { println!("🔧 [OpenAI] {} tools called:", tool_calls.len()); for tc in tool_calls { @@ -240,7 +252,7 @@ impl OpenAIProvider { } } } else { - println!("🔧 [OpenAI] No tools called"); + println!("🔧 [OpenAI] No tools called (no tool_calls field)"); } // Process tool calls if any @@ -318,26 +330,203 @@ impl OpenAIProvider { match function_name.as_str() { "get_memories" => { - - let _limit = arguments.get("limit").and_then(|v| v.as_i64()).unwrap_or(5); - // TODO: Implement actual MCP call - Ok(json!({"info": "記憶機能は実装中です"})) + let limit = arguments.get("limit").and_then(|v| v.as_i64()).unwrap_or(5); + + // MCP server call to get memories + match self.service_client.get_request(&format!("http://localhost:8080/memories/{}", context_user_id)).await { + Ok(result) => { + // Extract the actual memory content from MCP response + if let Some(content) = result.get("result").and_then(|r| r.get("content")) { + if let Some(text_content) = content.get(0).and_then(|c| c.get("text")) { + // Parse the text content as JSON (it's a serialized array) + if let Ok(memories_array) = serde_json::from_str::>(text_content.as_str().unwrap_or("[]")) { + let limited_memories: Vec = memories_array.into_iter().take(limit as usize).collect(); + Ok(json!({ + "memories": limited_memories, + "count": limited_memories.len() + })) + } else { + Ok(json!({ + "memories": [text_content.as_str().unwrap_or("No memories found")], + "count": 1 + })) + } + } else { + Ok(json!({"memories": [], "count": 0, "info": "No memories available"})) + } + } else { + Ok(json!({"memories": [], "count": 0, "info": "No response from memory service"})) + } + } + Err(e) => { + Ok(json!({"error": format!("Failed to retrieve memories: {}", e)})) + } + } } "search_memories" => { - let _keywords = arguments.get("keywords").and_then(|v| v.as_array()); - // TODO: Implement actual MCP call - Ok(json!({"info": "記憶検索機能は実装中です"})) + let keywords = arguments.get("keywords").and_then(|v| v.as_array()).unwrap_or(&vec![]).clone(); + + // Convert keywords to strings + let keyword_strings: Vec = keywords.iter() + .filter_map(|k| k.as_str().map(|s| s.to_string())) + .collect(); + + if keyword_strings.is_empty() { + return Ok(json!({"error": "No keywords provided for search"})); + } + + // MCP server call to search memories + let search_request = json!({ + "keywords": keyword_strings + }); + + match self.service_client.post_request( + &format!("http://localhost:8080/memories/{}/search", context_user_id), + &search_request + ).await { + Ok(result) => { + // Extract the actual memory content from MCP response + if let Some(content) = result.get("result").and_then(|r| r.get("content")) { + if let Some(text_content) = content.get(0).and_then(|c| c.get("text")) { + // Parse the search results + if let Ok(search_result) = serde_json::from_str::>(text_content.as_str().unwrap_or("[]")) { + let memory_contents: Vec = search_result.iter() + .filter_map(|item| item.get("content").and_then(|c| c.as_str().map(|s| s.to_string()))) + .collect(); + + Ok(json!({ + "memories": memory_contents, + "count": memory_contents.len(), + "keywords": keyword_strings + })) + } else { + Ok(json!({ + "memories": [], + "count": 0, + "info": format!("No memories found for keywords: {}", keyword_strings.join(", ")) + })) + } + } else { + Ok(json!({"memories": [], "count": 0, "info": "No search results available"})) + } + } else { + Ok(json!({"memories": [], "count": 0, "info": "No response from search service"})) + } + } + Err(e) => { + Ok(json!({"error": format!("Failed to search memories: {}", e)})) + } + } } "get_contextual_memories" => { - let _query = arguments.get("query").and_then(|v| v.as_str()).unwrap_or(""); - let _limit = arguments.get("limit").and_then(|v| v.as_i64()).unwrap_or(5); - // TODO: Implement actual MCP call - Ok(json!({"info": "文脈記憶機能は実装中です"})) + let query = arguments.get("query").and_then(|v| v.as_str()).unwrap_or(""); + let limit = arguments.get("limit").and_then(|v| v.as_i64()).unwrap_or(5); + + if query.is_empty() { + return Ok(json!({"error": "No query provided for contextual search"})); + } + + // MCP server call to get contextual memories + let contextual_request = json!({ + "query": query, + "limit": limit + }); + + match self.service_client.post_request( + &format!("http://localhost:8080/memories/{}/contextual", context_user_id), + &contextual_request + ).await { + Ok(result) => { + // Extract the actual memory content from MCP response + if let Some(content) = result.get("result").and_then(|r| r.get("content")) { + if let Some(text_content) = content.get(0).and_then(|c| c.get("text")) { + // Parse contextual search results + if text_content.as_str().unwrap_or("").contains("Found") { + // Extract memories from the formatted text response + let text = text_content.as_str().unwrap_or(""); + if let Some(json_start) = text.find('[') { + if let Ok(memories_result) = serde_json::from_str::>(&text[json_start..]) { + let memory_contents: Vec = memories_result.iter() + .filter_map(|item| item.get("content").and_then(|c| c.as_str().map(|s| s.to_string()))) + .collect(); + + Ok(json!({ + "memories": memory_contents, + "count": memory_contents.len(), + "query": query + })) + } else { + Ok(json!({ + "memories": [], + "count": 0, + "info": format!("No contextual memories found for: {}", query) + })) + } + } else { + Ok(json!({ + "memories": [], + "count": 0, + "info": format!("No contextual memories found for: {}", query) + })) + } + } else { + Ok(json!({ + "memories": [], + "count": 0, + "info": format!("No contextual memories found for: {}", query) + })) + } + } else { + Ok(json!({"memories": [], "count": 0, "info": "No contextual results available"})) + } + } else { + Ok(json!({"memories": [], "count": 0, "info": "No response from contextual search service"})) + } + } + Err(e) => { + Ok(json!({"error": format!("Failed to get contextual memories: {}", e)})) + } + } } "get_relationship" => { - let _user_id = arguments.get("user_id").and_then(|v| v.as_str()).unwrap_or(context_user_id); - // TODO: Implement actual MCP call - Ok(json!({"info": "関係性機能は実装中です"})) + let target_user_id = arguments.get("user_id").and_then(|v| v.as_str()).unwrap_or(context_user_id); + + // MCP server call to get relationship status + match self.service_client.get_request(&format!("http://localhost:8080/status/{}", target_user_id)).await { + Ok(result) => { + // Extract relationship information from MCP response + if let Some(content) = result.get("result").and_then(|r| r.get("content")) { + if let Some(text_content) = content.get(0).and_then(|c| c.get("text")) { + // Parse the status response to extract relationship data + if let Ok(status_data) = serde_json::from_str::(text_content.as_str().unwrap_or("{}")) { + if let Some(relationship) = status_data.get("relationship") { + Ok(json!({ + "relationship": relationship, + "user_id": target_user_id + })) + } else { + Ok(json!({ + "info": format!("No relationship found for user: {}", target_user_id), + "user_id": target_user_id + })) + } + } else { + Ok(json!({ + "info": format!("Could not parse relationship data for user: {}", target_user_id), + "user_id": target_user_id + })) + } + } else { + Ok(json!({"info": "No relationship data available", "user_id": target_user_id})) + } + } else { + Ok(json!({"info": "No response from relationship service", "user_id": target_user_id})) + } + } + Err(e) => { + Ok(json!({"error": format!("Failed to get relationship: {}", e)})) + } + } } // ai.card tools "card_get_user_cards" => { diff --git a/src/persona.rs b/src/persona.rs index 4f3549c..239b621 100644 --- a/src/persona.rs +++ b/src/persona.rs @@ -124,7 +124,13 @@ impl Persona { .map_err(|_| anyhow::anyhow!("OpenAI API key not found. Set OPENAI_API_KEY environment variable or add to config."))?; let openai_model = model.unwrap_or_else(|| "gpt-4".to_string()); - let openai_provider = OpenAIProvider::new(api_key, Some(openai_model)); + + // Get system prompt from config + let system_prompt = self.config.providers.get("openai") + .and_then(|p| p.system_prompt.clone()); + + + let openai_provider = OpenAIProvider::with_system_prompt(api_key, Some(openai_model), system_prompt); // Use OpenAI with MCP tools support openai_provider.chat_with_mcp(message.to_string(), user_id.to_string()).await?