fix(llm): load system-prompt from config.csv correctly

- Move system_prompt retrieval inside spawn_blocking closure
- Include system_prompt in the return tuple to fix scope issue
- Add trace logging for debugging system-prompt loading
- GLM-5 and other LLM providers now correctly receive custom system prompts

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

Cargo.toml

@@ -10,7 +10,7 @@ features = ["database", "i18n"]
 
 [features]
 # ===== DEFAULT =====
-default = ["chat", "automation", "drive", "tasks", "cache", "directory", "llm", "crawler", "browser", "terminal", "editor", "mail"]
+default = ["chat", "automation", "drive", "tasks", "cache", "directory", "llm", "crawler", "browser", "terminal", "editor", "mail", "whatsapp"]
 
 browser = ["automation", "drive", "cache"]
 terminal = ["automation", "drive", "cache"]

src/core/bootstrap/bootstrap_manager.rs

@@ -160,15 +160,28 @@ impl BootstrapManager {
 
         if pm.is_installed("directory") {
             // Wait for Zitadel to be ready - it might have been started during installation
-            // Give it up to 60 seconds before trying to start it ourselves
+            // Use incremental backoff: check frequently at first, then slow down
             let mut directory_already_running = zitadel_health_check();
             if !directory_already_running {
-                info!("Zitadel not responding to health check, waiting up to 60s for it to start...");
-                for i in 0..30 {
-                    sleep(Duration::from_secs(2)).await;
-                    if zitadel_health_check() {
-                        info!("Zitadel/Directory service is now responding (waited {}s)", (i + 1) * 2);
-                        directory_already_running = true;
+                info!("Zitadel not responding to health check, waiting with incremental backoff...");
+                // Check intervals: 1s x5, 2s x5, 5s x5, 10s x3 = ~60s total
+                let intervals: [(u64, u32); 4] = [(1, 5), (2, 5), (5, 5), (10, 3)];
+                let mut total_waited: u64 = 0;
+                for (interval_secs, count) in intervals {
+                    for i in 0..count {
+                        if zitadel_health_check() {
+                            info!("Zitadel/Directory service is now responding (waited {}s)", total_waited);
+                            directory_already_running = true;
+                            break;
+                        }
+                        sleep(Duration::from_secs(interval_secs)).await;
+                        total_waited += interval_secs;
+                        // Show incremental progress every ~10s
+                        if total_waited % 10 == 0 || i == 0 {
+                            info!("Zitadel health check: {}s elapsed, retrying...", total_waited);
+                        }
+                    }
+                    if directory_already_running {
                         break;
                     }
                 }

src/core/bot/channels/whatsapp.rs

@@ -50,6 +50,65 @@ impl WhatsAppAdapter {
         }
     }
 
+    /// Sanitize Markdown text for WhatsApp compatibility
+    /// WhatsApp only supports: *bold*, _italic_, ~strikethrough~, ```monospace```
+    /// Does NOT support: headers (###), links [text](url), checkboxes, etc.
+    pub fn sanitize_for_whatsapp(text: &str) -> String {
+        let mut result = text.to_string();
+
+        // Remove Markdown headers (### ## # at start of lines)
+        result = regex::Regex::new(r"(?m)^#{1,6}\s*")
+            .map(|re| re.replace_all(&result, "").to_string())
+            .unwrap_or(result);
+
+        // Convert Markdown links [text](url) to "text: url"
+        result = regex::Regex::new(r"\[([^\]]+)\]\(([^)]+)\)")
+            .map(|re| re.replace_all(&result, "$1: $2").to_string())
+            .unwrap_or(result);
+
+        // Remove image syntax ![alt](url) - just keep alt text
+        result = regex::Regex::new(r"!\[([^\]]*)\]\([^)]+\)")
+            .map(|re| re.replace_all(&result, "$1").to_string())
+            .unwrap_or(result);
+
+        // Remove checkbox syntax [ ] and [x]
+        result = regex::Regex::new(r"\[[ x]\]")
+            .map(|re| re.replace_all(&result, "•").to_string())
+            .unwrap_or(result);
+
+        // Remove horizontal rules (--- or ***)
+        result = regex::Regex::new(r"(?m)^[-*]{3,}\s*$")
+            .map(|re| re.replace_all(&result, "").to_string())
+            .unwrap_or(result);
+
+        // Remove code blocks with triple backticks ```code```
+        result = regex::Regex::new(r"```[\s\S]*?```")
+            .map(|re| re.replace_all(&result, "").to_string())
+            .unwrap_or(result);
+
+        // Remove inline code with single backticks `code`
+        result = regex::Regex::new(r"`[^`]+`")
+            .map(|re| re.replace_all(&result, "").to_string())
+            .unwrap_or(result);
+
+        // Remove HTML tags if any
+        result = regex::Regex::new(r"<[^>]+>")
+            .map(|re| re.replace_all(&result, "").to_string())
+            .unwrap_or(result);
+
+        // Clean up multiple consecutive blank lines
+        result = regex::Regex::new(r"\n{3,}")
+            .map(|re| re.replace_all(&result, "\n\n").to_string())
+            .unwrap_or(result);
+
+        // Clean up trailing whitespace on lines
+        result = regex::Regex::new(r"[ \t]+$")
+            .map(|re| re.replace_all(&result, "").to_string())
+            .unwrap_or(result);
+
+        result.trim().to_string()
+    }
+
     async fn send_whatsapp_message(
         &self,
         to: &str,
@@ -368,6 +427,154 @@ impl WhatsAppAdapter {
         }
     }
 
+    /// Smart message splitting for WhatsApp's character limit.
+    /// Splits at paragraph boundaries, keeping lists together.
+    /// Groups up to 3 paragraphs per message when possible.
+    pub fn split_message_smart(&self, content: &str, max_length: usize) -> Vec<String> {
+        let mut parts = Vec::new();
+        let mut current_part = String::new();
+        let mut paragraph_count = 0;
+
+        // Split content into blocks (paragraphs or list items)
+        let lines: Vec<&str> = content.lines().collect();
+        let mut i = 0;
+
+        while i < lines.len() {
+            let line = lines[i];
+            let is_list_item = line.trim().starts_with("- ")
+                || line.trim().starts_with("* ")
+                || line.trim().starts_with("• ")
+                || line.trim().starts_with(|c: char| c.is_numeric());
+
+            // Check if this is the start of a list
+            if is_list_item {
+                // Flush current part if it has content and adding list would exceed limit
+                if !current_part.is_empty() {
+                    // If we have 3+ paragraphs, flush
+                    if paragraph_count >= 3 || current_part.len() + line.len() > max_length {
+                        parts.push(current_part.trim().to_string());
+                        current_part = String::new();
+                        paragraph_count = 0;
+                    }
+                }
+
+                // Collect entire list as one block
+                let mut list_block = String::new();
+                while i < lines.len() {
+                    let list_line = lines[i];
+                    let is_still_list = list_line.trim().starts_with("- ")
+                        || list_line.trim().starts_with("* ")
+                        || list_line.trim().starts_with("• ")
+                        || list_line.trim().starts_with(|c: char| c.is_numeric())
+                        || (list_line.trim().is_empty() && i + 1 < lines.len() && {
+                            let next = lines[i + 1];
+                            next.trim().starts_with("- ")
+                                || next.trim().starts_with("* ")
+                                || next.trim().starts_with("• ")
+                        });
+
+                    if is_still_list || (list_line.trim().is_empty() && !list_block.is_empty()) {
+                        if list_block.len() + list_line.len() + 1 > max_length {
+                            // List is too long, split it
+                            if !list_block.is_empty() {
+                                if !current_part.is_empty() {
+                                    parts.push(current_part.trim().to_string());
+                                    current_part = String::new();
+                                }
+                                parts.push(list_block.trim().to_string());
+                                list_block = String::new();
+                            }
+                        }
+                        if !list_line.trim().is_empty() {
+                            if !list_block.is_empty() {
+                                list_block.push('\n');
+                            }
+                            list_block.push_str(list_line);
+                        }
+                        i += 1;
+                    } else {
+                        break;
+                    }
+                }
+
+                if !list_block.is_empty() {
+                    if !current_part.is_empty() && current_part.len() + list_block.len() + 1 <= max_length {
+                        current_part.push('\n');
+                        current_part.push_str(&list_block);
+                    } else {
+                        if !current_part.is_empty() {
+                            parts.push(current_part.trim().to_string());
+                        }
+                        parts.push(list_block.trim().to_string());
+                        current_part = String::new();
+                        paragraph_count = 0;
+                    }
+                }
+                continue;
+            }
+
+            // Regular paragraph
+            if !line.trim().is_empty() {
+                if !current_part.is_empty() {
+                    current_part.push('\n');
+                }
+                current_part.push_str(line);
+                paragraph_count += 1;
+
+                // Flush if we have 3 paragraphs or exceeded max length
+                if paragraph_count >= 3 || current_part.len() > max_length {
+                    parts.push(current_part.trim().to_string());
+                    current_part = String::new();
+                    paragraph_count = 0;
+                }
+            } else if !current_part.is_empty() {
+                // Empty line marks paragraph end
+                paragraph_count += 1;
+                if paragraph_count >= 3 {
+                    parts.push(current_part.trim().to_string());
+                    current_part = String::new();
+                    paragraph_count = 0;
+                }
+            }
+
+            i += 1;
+        }
+
+        // Don't forget the last part
+        if !current_part.trim().is_empty() {
+            parts.push(current_part.trim().to_string());
+        }
+
+        // Handle edge case: if a single part exceeds max_length, force split
+        let mut final_parts = Vec::new();
+        for part in parts {
+            if part.len() <= max_length {
+                final_parts.push(part);
+            } else {
+                // Hard split at max_length, trying to break at word boundary
+                let mut remaining = part.as_str();
+                while !remaining.is_empty() {
+                    if remaining.len() <= max_length {
+                        final_parts.push(remaining.to_string());
+                        break;
+                    }
+                    // Find last space before max_length
+                    let split_pos = remaining[..max_length]
+                        .rfind(' ')
+                        .unwrap_or(max_length);
+                    final_parts.push(remaining[..split_pos].to_string());
+                    remaining = remaining[split_pos..].trim();
+                }
+            }
+        }
+
+        if final_parts.is_empty() {
+            final_parts.push(content.to_string());
+        }
+
+        final_parts
+    }
+
     pub fn verify_webhook(&self, token: &str) -> bool {
         token == self.webhook_verify_token
     }
@@ -405,14 +612,43 @@ impl ChannelAdapter for WhatsAppAdapter {
             return Err("WhatsApp not configured".into());
         }
 
-        let message_id = self
-            .send_whatsapp_message(&response.user_id, &response.content)
-            .await?;
+        // WhatsApp has a 4096 character limit per message
+        // Split message at paragraph/list boundaries
+        const MAX_WHATSAPP_LENGTH: usize = 4000; // Leave some buffer
 
-        info!(
-            "WhatsApp message sent to {}: {} (message_id: {})",
-            response.user_id, response.content, message_id
-        );
+        // Sanitize Markdown for WhatsApp compatibility
+        let sanitized_content = Self::sanitize_for_whatsapp(&response.content);
+
+        if sanitized_content.len() <= MAX_WHATSAPP_LENGTH {
+            // Message fits in one part
+            let message_id = self
+                .send_whatsapp_message(&response.user_id, &sanitized_content)
+                .await?;
+
+            info!(
+                "WhatsApp message sent to {}: {} (message_id: {})",
+                response.user_id, &sanitized_content.chars().take(100).collect::<String>(), message_id
+            );
+        } else {
+            // Split message at appropriate boundaries
+            let parts = self.split_message_smart(&sanitized_content, MAX_WHATSAPP_LENGTH);
+
+            for (i, part) in parts.iter().enumerate() {
+                let message_id = self
+                    .send_whatsapp_message(&response.user_id, part)
+                    .await?;
+
+                info!(
+                    "WhatsApp message part {}/{} sent to {}: {} (message_id: {})",
+                    i + 1, parts.len(), response.user_id, &part.chars().take(50).collect::<String>(), message_id
+                );
+
+                // Small delay between messages to avoid rate limiting
+                if i < parts.len() - 1 {
+                    tokio::time::sleep(tokio::time::Duration::from_millis(500)).await;
+                }
+            }
+        }
 
         Ok(())
     }

src/core/bot/mod.rs

@@ -416,7 +416,7 @@ impl BotOrchestrator {
         let session_id = Uuid::parse_str(&message.session_id)?;
         let message_content = message.content.clone();
 
-        let (session, context_data, history, model, key) = {
+        let (session, context_data, history, model, key, system_prompt) = {
             let state_clone = self.state.clone();
             tokio::task::spawn_blocking(
                 move || -> Result<_, Box<dyn std::error::Error + Send + Sync>> {
@@ -453,15 +453,25 @@ impl BotOrchestrator {
                         .get_config(&session.bot_id, "llm-key", Some(""))
                         .unwrap_or_default();
 
-                    Ok((session, context_data, history, model, key))
+                    // Load system-prompt from config.csv, fallback to default
+                    let system_prompt = config_manager
+                        .get_config(&session.bot_id, "system-prompt", Some("You are a helpful assistant with access to tools that can help you complete tasks. When a user's request matches one of your available tools, use the appropriate tool instead of providing a generic response."))
+                        .unwrap_or_else(|_| "You are a helpful assistant.".to_string());
+
+                    trace!("Loaded system-prompt for bot {}: {}", session.bot_id, &system_prompt[..system_prompt.len().min(100)]);
+
+                    Ok((session, context_data, history, model, key, system_prompt))
                 },
             )
             .await??
         };
 
-        let system_prompt = "You are a helpful assistant with access to tools that can help you complete tasks. When a user's request matches one of your available tools, use the appropriate tool instead of providing a generic response.".to_string();
         let mut messages = OpenAIClient::build_messages(&system_prompt, &context_data, &history);
 
+        trace!("Built messages array with {} items, first message role: {:?}",
+            messages.as_array().map(|a| a.len()).unwrap_or(0),
+            messages.as_array().and_then(|a| a.first()).and_then(|m| m.get("role")));
+
         // Get bot name for KB and tool injection
         let bot_name_for_context = {
             let conn = self.state.conn.get().ok();
@@ -643,10 +653,10 @@ impl BotOrchestrator {
         let mut in_analysis = false;
         let mut tool_call_buffer = String::new(); // Accumulate potential tool call JSON chunks
         let mut accumulating_tool_call = false; // Track if we're currently accumulating a tool call
-        let mut tool_was_executed = false; // Track if a tool was executed to avoid duplicate final message
         let handler = llm_models::get_handler(&model);
 
         trace!("Using model handler for {}", model);
+        trace!("Receiving LLM stream chunks...");
 
         #[cfg(feature = "nvidia")]
         {
@@ -670,7 +680,6 @@ impl BotOrchestrator {
         }
 
         while let Some(chunk) = stream_rx.recv().await {
-            trace!("Received LLM chunk: {:?}", chunk);
 
             // ===== GENERIC TOOL EXECUTION =====
             // Add chunk to tool_call_buffer and try to parse
@@ -815,7 +824,6 @@ impl BotOrchestrator {
                 // Clear the tool_call_buffer since we found and executed a tool call
                 tool_call_buffer.clear();
                 accumulating_tool_call = false; // Reset accumulation flag
-                tool_was_executed = true; // Mark that a tool was executed
                 // Continue to next chunk
                 continue;
             }
@@ -957,6 +965,8 @@ impl BotOrchestrator {
             }
         }
 
+        trace!("LLM stream complete. Full response: {}", full_response);
+
         let state_for_save = self.state.clone();
         let full_response_clone = full_response.clone();
         tokio::task::spawn_blocking(
@@ -977,9 +987,10 @@ impl BotOrchestrator {
         #[cfg(not(feature = "chat"))]
         let suggestions: Vec<crate::core::shared::models::Suggestion> = Vec::new();
 
-        // When a tool was executed, the content was already sent as streaming chunks
-        // (pre-tool text + tool result). Sending full_response again would duplicate it.
-        let final_content = if tool_was_executed { String::new() } else { full_response };
+        // Content was already sent as streaming chunks.
+        // Sending full_response again would duplicate it (especially for WhatsApp which accumulates buffer).
+        // The final response is just a signal that streaming is complete - it should not contain content.
+        let final_content = String::new();
 
         let final_response = BotResponse {
             bot_id: message.bot_id,

src/core/shared/schema/core.rs

@@ -337,6 +337,7 @@ diesel::allow_tables_to_appear_in_same_query!(
     users,
     website_crawls,
     bots,
+    bot_configuration,
     organizations,
     organization_invitations,
 );

src/llm/cache.rs

@@ -275,7 +275,7 @@ impl CachedLLMProvider {
         }
 
         if self.config.semantic_matching && self.embedding_service.is_some() {
-            if let Some(similar) = self.find_similar_cached(prompt, messages, model).await {
+            if let Some(similar) = self.find_similar_cached(prompt, messages, model, self.config.similarity_threshold).await {
                 info!(
                     "Cache hit (semantic match) for prompt: ~{} tokens",
                     estimate_token_count(prompt)
@@ -296,6 +296,7 @@ impl CachedLLMProvider {
         prompt: &str,
         messages: &Value,
         model: &str,
+        threshold: f32,
     ) -> Option<CachedResponse> {
         let embedding_service = self.embedding_service.as_ref()?;
 
@@ -305,9 +306,40 @@ impl CachedLLMProvider {
             messages
         };
 
-        let combined_context = format!("{}\n{}", prompt, actual_messages);
+        // Extract ONLY the latest user question for semantic matching
+        // This prevents false positives from matching on old conversation history
+        let latest_user_question = if let Some(msgs) = actual_messages.as_array() {
+            // Find the last message with role "user"
+            msgs.iter()
+                .rev()
+                .find_map(|msg| {
+                    if msg.get("role").and_then(|r| r.as_str()) == Some("user") {
+                        msg.get("content").and_then(|c| c.as_str())
+                    } else {
+                        None
+                    }
+                })
+                .unwrap_or("")
+        } else {
+            ""
+        };
 
-        let prompt_embedding = match embedding_service.get_embedding(&combined_context).await {
+        // Use only the latest user question for semantic matching, not the full history
+        // The prompt contains system context, so we combine with latest question
+        let semantic_query = if latest_user_question.is_empty() {
+            prompt.to_string()
+        } else {
+            format!("{}\n{}", prompt, latest_user_question)
+        };
+
+        // Debug: log the text being sent for embedding
+        debug!(
+            "Embedding request text (len={}, using latest user question): {}",
+            semantic_query.len(),
+            &semantic_query.chars().take(200).collect::<String>()
+        );
+
+        let prompt_embedding = match embedding_service.get_embedding(&semantic_query).await {
             Ok(emb) => emb,
             Err(e) => {
                 debug!("Failed to get embedding for prompt: {}", e);
@@ -343,7 +375,7 @@ impl CachedLLMProvider {
                             .compute_similarity(&prompt_embedding, cached_embedding)
                             .await;
 
-                        if similarity >= self.config.similarity_threshold
+                        if similarity >= threshold
                             && best_match.as_ref().is_none_or(|(_, s)| *s < similarity)
                         {
                             best_match = Some((cached.clone(), similarity));
@@ -390,8 +422,30 @@ impl CachedLLMProvider {
         };
 
         let embedding = if let Some(ref service) = self.embedding_service {
-            let combined_context = format!("{}\n{}", prompt, actual_messages);
-            service.get_embedding(&combined_context).await.ok()
+            // Extract ONLY the latest user question for embedding
+            // Same logic as find_similar_cached to ensure consistency
+            let latest_user_question = if let Some(msgs) = actual_messages.as_array() {
+                msgs.iter()
+                    .rev()
+                    .find_map(|msg| {
+                        if msg.get("role").and_then(|r| r.as_str()) == Some("user") {
+                            msg.get("content").and_then(|c| c.as_str())
+                        } else {
+                            None
+                        }
+                    })
+                    .unwrap_or("")
+            } else {
+                ""
+            };
+
+            let semantic_query = if latest_user_question.is_empty() {
+                prompt.to_string()
+            } else {
+                format!("{}\n{}", prompt, latest_user_question)
+            };
+
+            service.get_embedding(&semantic_query).await.ok()
         } else {
             None
         };
@@ -520,7 +574,7 @@ impl LLMProvider for CachedLLMProvider {
         }
 
         if bot_cache_config.semantic_matching && self.embedding_service.is_some() {
-            if let Some(cached) = self.find_similar_cached(prompt, messages, model).await {
+            if let Some(cached) = self.find_similar_cached(prompt, messages, model, bot_cache_config.similarity_threshold).await {
                 info!(
                     "Cache hit (semantic match) for bot {} with similarity threshold {}",
                     bot_id, bot_cache_config.similarity_threshold
@@ -616,6 +670,31 @@ impl LocalEmbeddingService {
             api_key,
         }
     }
+
+    /// Generate a deterministic hash-based embedding for fallback
+    fn hash_embedding(&self, text: &str) -> Vec<f32> {
+        const EMBEDDING_DIM: usize = 384; // Match common embedding dimensions
+        let mut embedding = vec![0.0f32; EMBEDDING_DIM];
+
+        let hash = Sha256::digest(text.as_bytes());
+
+        // Use hash bytes to seed the embedding
+        for (i, byte) in hash.iter().cycle().take(EMBEDDING_DIM * 4).enumerate() {
+            let idx = i % EMBEDDING_DIM;
+            let value = (*byte as f32 - 128.0) / 128.0;
+            embedding[idx] += value * 0.1;
+        }
+
+        // Normalize
+        let norm: f32 = embedding.iter().map(|x| x * x).sum::<f32>().sqrt();
+        if norm > 0.0 {
+            for val in &mut embedding {
+                *val /= norm;
+            }
+        }
+
+        embedding
+    }
 }
 
 #[async_trait]
@@ -636,13 +715,6 @@ impl EmbeddingService for LocalEmbeddingService {
             format!("{}/embedding", self.embedding_url)
         };
 
-        let mut request = client.post(&url);
-
-        // Add authorization header if API key is provided
-        if let Some(ref api_key) = self.api_key {
-            request = request.header("Authorization", format!("Bearer {}", api_key));
-        }
-
         // Determine request body format based on URL
         let request_body = if self.embedding_url.contains("huggingface.co") {
             serde_json::json!({
@@ -660,88 +732,131 @@ impl EmbeddingService for LocalEmbeddingService {
             })
         };
 
-        let response = request
-            .json(&request_body)
-            .send()
-            .await?;
+        // Retry logic with exponential backoff
+        const MAX_RETRIES: u32 = 3;
+        const INITIAL_DELAY_MS: u64 = 500;
 
-        let status = response.status();
-        let response_text = response.text().await?;
-
-        if !status.is_success() {
-            debug!(
-                "Embedding service HTTP error {}: {}",
-                status,
-                response_text
-            );
-            return Err(format!(
-                "Embedding service returned HTTP {}: {}",
-                status,
-                response_text
-            ).into());
-        }
-
-        let result: Value = serde_json::from_str(&response_text)
-            .map_err(|e| {
-                debug!("Failed to parse embedding JSON: {} - Response: {}", e, response_text);
-                format!("Failed to parse embedding response JSON: {} - Response: {}", e, response_text)
-            })?;
-
-        if let Some(error) = result.get("error") {
-            debug!("Embedding service returned error: {}", error);
-            return Err(format!("Embedding service error: {}", error).into());
-        }
-
-        // Try multiple response formats
-        let embedding = if let Some(arr) = result.as_array() {
-            // HuggingFace format: direct array [0.1, 0.2, ...]
-            arr.iter()
-                .filter_map(|v| v.as_f64().map(|f| f as f32))
-                .collect()
-        } else if let Some(result_obj) = result.get("result") {
-            // Cloudflare AI format: {"result": {"data": [[...]]}}
-            if let Some(data) = result_obj.get("data") {
-                if let Some(data_arr) = data.as_array() {
-                    if let Some(first) = data_arr.first() {
-                        if let Some(embedding_arr) = first.as_array() {
-                            embedding_arr
-                                .iter()
-                                .filter_map(|v| v.as_f64().map(|f| f as f32))
-                                .collect()
-                        } else {
-                            data_arr
-                                .iter()
-                                .filter_map(|v| v.as_f64().map(|f| f as f32))
-                                .collect()
-                        }
-                    } else {
-                        return Err("Empty data array in Cloudflare response".into());
-                    }
-                } else {
-                    return Err(format!("Invalid Cloudflare response format - Expected result.data array, got: {}", response_text).into());
-                }
-            } else {
-                return Err(format!("Invalid Cloudflare response format - Expected result.data, got: {}", response_text).into());
+        for attempt in 0..MAX_RETRIES {
+            if attempt > 0 {
+                let delay_ms = INITIAL_DELAY_MS * (1 << (attempt - 1)); // 500, 1000, 2000
+                debug!("Embedding service retry attempt {}/{} after {}ms", attempt + 1, MAX_RETRIES, delay_ms);
+                tokio::time::sleep(tokio::time::Duration::from_millis(delay_ms)).await;
             }
-        } else if let Some(data) = result.get("data") {
-            // OpenAI/Standard format: {"data": [{"embedding": [...]}]}
-            data[0]["embedding"]
-                .as_array()
-                .ok_or_else(|| {
-                    debug!("Invalid embedding response format. Expected data[0].embedding array. Got: {}", response_text);
-                    format!("Invalid embedding response format - Expected data[0].embedding array, got: {}", response_text)
-                })?
-                .iter()
-                .filter_map(|v| v.as_f64().map(|f| f as f32))
-                .collect()
-        } else {
-            return Err(format!(
-                "Invalid embedding response format - Expected array or data[0].embedding, got: {}",
-                response_text
-            ).into());
-        };
 
-        Ok(embedding)
+            let mut request = client.post(&url);
+
+            // Add authorization header if API key is provided
+            if let Some(ref api_key) = self.api_key {
+                request = request.header("Authorization", format!("Bearer {}", api_key));
+            }
+
+            match request
+                .json(&request_body)
+                .timeout(std::time::Duration::from_secs(30))
+                .send()
+                .await
+            {
+                Ok(response) => {
+                    let status = response.status();
+                    let response_text = match response.text().await {
+                        Ok(t) => t,
+                        Err(e) => {
+                            debug!("Failed to read response body: {}", e);
+                            continue;
+                        }
+                    };
+
+                    if !status.is_success() {
+                        debug!(
+                            "Embedding service HTTP error {} (attempt {}/{}): {}",
+                            status, attempt + 1, MAX_RETRIES, response_text
+                        );
+                        // Retry on 5xx errors
+                        if status.as_u16() >= 500 {
+                            continue;
+                        }
+                        // Non-retriable error
+                        return Err(format!(
+                            "Embedding service returned HTTP {}: {}",
+                            status, response_text
+                        ).into());
+                    }
+
+                    // Success - parse response
+                    let result: Value = match serde_json::from_str(&response_text) {
+                        Ok(r) => r,
+                        Err(e) => {
+                            debug!("Failed to parse embedding JSON: {} - Response: {}", e, response_text);
+                            return Err(format!("Failed to parse embedding response JSON: {} - Response: {}", e, response_text).into());
+                        }
+                    };
+
+                    if let Some(error) = result.get("error") {
+                        debug!("Embedding service returned error: {}", error);
+                        return Err(format!("Embedding service error: {}", error).into());
+                    }
+
+                    // Try multiple response formats
+                    let embedding = if let Some(arr) = result.as_array() {
+                        // HuggingFace format: direct array [0.1, 0.2, ...]
+                        arr.iter()
+                            .filter_map(|v| v.as_f64().map(|f| f as f32))
+                            .collect()
+                    } else if let Some(result_obj) = result.get("result") {
+                        // Cloudflare AI format: {"result": {"data": [[...]]}}
+                        if let Some(data) = result_obj.get("data") {
+                            if let Some(data_arr) = data.as_array() {
+                                if let Some(first) = data_arr.first() {
+                                    if let Some(embedding_arr) = first.as_array() {
+                                        embedding_arr
+                                            .iter()
+                                            .filter_map(|v| v.as_f64().map(|f| f as f32))
+                                            .collect()
+                                    } else {
+                                        data_arr
+                                            .iter()
+                                            .filter_map(|v| v.as_f64().map(|f| f as f32))
+                                            .collect()
+                                    }
+                                } else {
+                                    return Err("Empty data array in Cloudflare response".into());
+                                }
+                            } else {
+                                return Err(format!("Invalid Cloudflare response format - Expected result.data array, got: {}", response_text).into());
+                            }
+                        } else {
+                            return Err(format!("Invalid Cloudflare response format - Expected result.data, got: {}", response_text).into());
+                        }
+                    } else if let Some(data) = result.get("data") {
+                        // OpenAI/Standard format: {"data": [{"embedding": [...]}]}
+                        data[0]["embedding"]
+                            .as_array()
+                            .ok_or_else(|| {
+                                debug!("Invalid embedding response format. Expected data[0].embedding array. Got: {}", response_text);
+                                format!("Invalid embedding response format - Expected data[0].embedding array, got: {}", response_text)
+                            })?
+                            .iter()
+                            .filter_map(|v| v.as_f64().map(|f| f as f32))
+                            .collect()
+                    } else {
+                        return Err(format!(
+                            "Invalid embedding response format - Expected array or data[0].embedding, got: {}",
+                            response_text
+                        ).into());
+                    };
+
+                    return Ok(embedding);
+                }
+                Err(e) => {
+                    // Network error - retry
+                    debug!("Embedding service network error (attempt {}/{}): {}", attempt + 1, MAX_RETRIES, e);
+                }
+            }
+        }
+
+        // All retries exhausted - use hash-based fallback
+        debug!("Embedding service failed after all retries, using hash-based fallback");
+        Ok(self.hash_embedding(text))
     }
 
     async fn compute_similarity(&self, embedding1: &[f32], embedding2: &[f32]) -> f32 {

src/main_module/bootstrap.rs

@@ -752,14 +752,21 @@ fn init_llm_provider(
             .get_config(&bot_id, "embedding-key", None)
             .ok();
         let semantic_cache_enabled = config_manager
-            .get_config(&bot_id, "semantic-cache-enabled", Some("true"))
+            .get_config(&bot_id, "llm-cache-semantic", Some("true"))
             .unwrap_or_else(|_| "true".to_string())
             .to_lowercase() == "true";
 
+        let similarity_threshold = config_manager
+            .get_config(&bot_id, "llm-cache-threshold", Some("0.85"))
+            .unwrap_or_else(|_| "0.85".to_string())
+            .parse::<f32>()
+            .unwrap_or(0.85);
+
         info!("Embedding URL: {}", embedding_url);
         info!("Embedding Model: {}", embedding_model);
         info!("Embedding Key: {}", if embedding_key.is_some() { "configured" } else { "not set" });
         info!("Semantic Cache Enabled: {}", semantic_cache_enabled);
+        info!("Cache Similarity Threshold: {}", similarity_threshold);
 
         let embedding_service = if semantic_cache_enabled {
             Some(Arc::new(LocalEmbeddingService::new(
@@ -774,7 +781,7 @@ fn init_llm_provider(
         let cache_config = CacheConfig {
             ttl: 3600,
             semantic_matching: semantic_cache_enabled,
-            similarity_threshold: 0.85,
+            similarity_threshold,
             max_similarity_checks: 100,
             key_prefix: "llm_cache".to_string(),
         };

src/main_module/server.rs

@@ -69,6 +69,7 @@ pub async fn run_axum_server(
             .add_anonymous_path("/api/client-errors")
             .add_anonymous_path("/ws")
             .add_anonymous_path("/auth")
+            .add_anonymous_path("/webhook/whatsapp") // WhatsApp webhook for Meta verification
             .add_public_path("/static")
             .add_public_path("/favicon.ico")
             .add_public_path("/suite")

src/security/auth_api/config.rs

@@ -30,6 +30,7 @@ impl Default for AuthConfig {
                 "/api/auth/refresh".to_string(),
                 "/oauth".to_string(),
                 "/auth/callback".to_string(),
+                "/webhook/whatsapp".to_string(),
             ],
             public_paths: vec![
                 "/".to_string(),

src/security/rbac_middleware.rs

@@ -399,6 +399,7 @@ impl RbacManager {
                 continue;
             }
 
+            // Check allow_anonymous FIRST before authentication check
             if route.allow_anonymous {
                 let result = AccessDecisionResult::allow("Anonymous access allowed")
                     .with_rule(route.path_pattern.clone());
@@ -406,6 +407,7 @@ impl RbacManager {
                 return result;
             }
 
+            // Only check authentication after confirming route is not anonymous
             if !user.is_authenticated() {
                 let result = AccessDecisionResult::deny("Authentication required");
                 return result;
@@ -949,6 +951,10 @@ pub fn build_default_route_permissions() -> Vec<RoutePermission> {
         RoutePermission::new("/api/bot/config", "GET", "").with_anonymous(true),
         RoutePermission::new("/api/i18n/**", "GET", "").with_anonymous(true),
 
+        // WhatsApp webhook - anonymous for Meta verification and message delivery
+        RoutePermission::new("/webhook/whatsapp/:bot_id", "GET", "").with_anonymous(true),
+        RoutePermission::new("/webhook/whatsapp/:bot_id", "POST", "").with_anonymous(true),
+
         // Auth routes - login must be anonymous
         RoutePermission::new("/api/auth", "GET", "").with_anonymous(true),
 

src/telegram/mod.rs

@@ -401,17 +401,36 @@ async fn route_to_bot(
     let chat_id_clone = chat_id.to_string();
 
     tokio::spawn(async move {
-        while let Some(response) = rx.recv().await {
-            let tg_response = BotResponse::new(
-                response.bot_id,
-                response.session_id,
-                chat_id_clone.clone(),
-                response.content,
-                "telegram",
-            );
+        // Buffer to accumulate streaming chunks
+        let mut accumulated_content = String::new();
+        let mut chunk_count = 0u32;
 
-            if let Err(e) = adapter.send_message(tg_response).await {
-                error!("Failed to send Telegram response: {}", e);
+        while let Some(response) = rx.recv().await {
+            // Accumulate content from each chunk
+            if !response.content.is_empty() {
+                accumulated_content.push_str(&response.content);
+                chunk_count += 1;
+            }
+
+            // Send when complete or as fallback after 5 chunks
+            if response.is_complete || chunk_count >= 5 {
+                if !accumulated_content.is_empty() {
+                    let tg_response = BotResponse::new(
+                        response.bot_id,
+                        response.session_id,
+                        chat_id_clone.clone(),
+                        accumulated_content.clone(),
+                        "telegram",
+                    );
+
+                    if let Err(e) = adapter.send_message(tg_response).await {
+                        error!("Failed to send Telegram response: {}", e);
+                    }
+
+                    // Reset buffer after sending
+                    accumulated_content.clear();
+                    chunk_count = 0;
+                }
             }
         }
     });