@retrivora-ai/rag-engine 2.0.7 → 2.1.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "2.0.7",
+  "version": "2.1.1",
   "description": "Retrivora AI is a plug-and-play AI engine for RAG chat experiences — generic vector DB + LLM provider, embeddable or standalone.",
   "author": "Abhinav Alkuchi",
   "license": "UNLICENSED",

package/src/components/ChatWindow.tsx

@@ -52,10 +52,10 @@ interface WindowWithSpeech extends Window {
   webkitSpeechRecognition?: new () => ISpeechRecognition;
 }
 
-export function ChatWindow({ 
-  className = '', 
-  style, 
-  onClose, 
+export function ChatWindow({
+  className = '',
+  style,
+  onClose,
   showClose = false,
   onResizeStart,
   onResetResize,
@@ -129,7 +129,7 @@ export function ChatWindow({
       alert('Speech recognition is not supported in your browser.');
       return;
     }
-    
+
     if (isListening) {
       recognitionRef.current.stop();
     } else {
@@ -178,7 +178,7 @@ export function ChatWindow({
   const sendMessage = useCallback(async () => {
     const text = input.trim();
     if (!text || isLoading) return;
-    
+
     if (isListening) {
       recognitionRef.current?.stop();
     }
@@ -250,11 +250,10 @@ export function ChatWindow({
   return (
     <div
       ref={windowRef}
-      className={`relative flex flex-col h-full flex-1 border border-slate-200 dark:border-white/10 shadow-2xl transition-all duration-300 ${currentRadius} ${
-        isGlass 
-          ? 'bg-white/90 dark:bg-[#0f0f1a]/90 backdrop-blur-xl' 
+      className={`relative flex flex-col h-full flex-1 border border-slate-200 dark:border-white/10 shadow-2xl transition-all duration-300 ${currentRadius} ${isGlass
+          ? 'bg-white/90 dark:bg-[#0f0f1a]/90 backdrop-blur-xl'
           : 'bg-white dark:bg-[#0f0f1a]'
-      } ${className}`}
+        } ${className}`}
       style={{ '--primary': ui.primaryColor, '--accent': ui.accentColor, ...style } as React.CSSProperties}
     >
       {/* Resize Handle (Top-Left) */}
@@ -293,11 +292,13 @@ export function ChatWindow({
         </div>
 
         <div className="flex items-center gap-1.5">
-          {/* Online indicator */}
-          <span className="flex items-center gap-1 text-[10px] text-emerald-500 dark:text-emerald-400 mr-2">
-            <span className="w-1.5 h-1.5 rounded-full bg-emerald-500 dark:bg-emerald-400 animate-pulse" />
-            Online
-          </span>
+          {/* Online indicator & Deployment ID badge */}
+          <div className="flex items-center gap-1.5 mr-2">
+            <span className="flex items-center gap-1 text-[10px] text-emerald-500 dark:text-emerald-400">
+              <span className="w-1.5 h-1.5 rounded-full bg-emerald-500 dark:bg-emerald-400 animate-pulse" />
+              Online
+            </span>
+          </div>
 
           <div className="flex items-center bg-slate-100/50 dark:bg-white/5 rounded-lg p-0.5 border border-slate-200/50 dark:border-white/5">
             {mounted && messages.length > 0 && (
@@ -427,9 +428,9 @@ export function ChatWindow({
             >
               <X className="w-5 h-5" />
             </button>
-            <DocumentUpload 
-              namespace={projectId} 
-              onUploadComplete={() => setIsUploadModalOpen(false)} 
+            <DocumentUpload
+              namespace={projectId}
+              onUploadComplete={() => setIsUploadModalOpen(false)}
             />
           </div>
         </div>
@@ -437,7 +438,7 @@ export function ChatWindow({
 
       {/* ── Input ── */}
       <div className={`px-4 pb-4 pt-2 border-t border-slate-200 dark:border-white/10 ${isGlass ? 'bg-transparent' : 'bg-white dark:bg-[#0f0f1a]'}`}>
-        
+
         {/* Dynamic Suggestions */}
         {(suggestions.length > 0 || isSuggesting) && !isLoading && (
           <div className="mb-3 flex flex-wrap gap-2 animate-in fade-in slide-in-from-bottom-2 duration-300">
@@ -488,11 +489,10 @@ export function ChatWindow({
               type="button"
               onClick={toggleListening}
               disabled={isLoading}
-              className={`flex-shrink-0 w-9 h-9 rounded-xl flex items-center justify-center transition-all hover:scale-105 active:scale-95 disabled:opacity-50 disabled:cursor-not-allowed disabled:hover:scale-100 disabled:active:scale-100 ${
-                isListening
+              className={`flex-shrink-0 w-9 h-9 rounded-xl flex items-center justify-center transition-all hover:scale-105 active:scale-95 disabled:opacity-50 disabled:cursor-not-allowed disabled:hover:scale-100 disabled:active:scale-100 ${isListening
                   ? 'bg-rose-500 text-white animate-pulse shadow-md shadow-rose-500/20'
                   : 'text-slate-400 dark:text-white/40 hover:bg-slate-200 dark:hover:bg-white/10'
-              }`}
+                }`}
               title={isListening ? 'Stop listening' : 'Start voice input'}
             >
               {isListening ? <MicOff className="w-4 h-4" /> : <Mic className="w-4 h-4" />}

package/src/core/Pipeline.ts

@@ -580,8 +580,8 @@ You are a helpful assistant. Use the provided context to answer questions accura
 
       const hasMetadataFilter = filter.metadata && Object.keys(filter.metadata).length > 0;
       const hasNumericPredicates = Array.isArray(filter.__numericPredicates) && filter.__numericPredicates.length > 0;
-      const wantsExhaustiveList = hasNumericPredicates || hasMetadataFilter || /\b(list|all|show|provide|give|get|browse|find|search|display)\b/i.test(question);
-      const retrievalLimit = wantsExhaustiveList ? Math.max(topK * 20, 100) : topK * 2;
+      const wantsExhaustiveList = true; // Always query all available resources across the database
+      const retrievalLimit = Math.max(topK * 50, 1000);
 
       const rawSources = ((strategyResult === 'vector' || strategyResult === 'both') && queryVector && queryVector.length > 0
         ? await this.vectorDB.query(queryVector, retrievalLimit, ns, filter)
@@ -602,52 +602,37 @@ You are a helpful assistant. Use the provided context to answer questions accura
       const embedMs = retrieveEnd - embedStart;
       const retrieveMs = retrieveEnd - embedStart;
 
-      // 3. Reranking
+      // 3. Reranking & Filtering across ALL matching resources
       const rerankStart = performance.now();
       const structuredSources = this.applyStructuredFilters(rawSources, filter).filter((s): s is VectorMatch => Boolean(s && typeof s === 'object'));
       let fullSources = hasNumericPredicates
         ? structuredSources
         : structuredSources.filter((m) => (m?.score ?? 0) >= scoreThreshold);
-      const rerankLimit = wantsExhaustiveList ? retrievalLimit : topK;
+      const rerankLimit = Math.max(retrievalLimit, fullSources.length);
       const useReranking = arch !== 'naive' && (arch === 'retrieve-and-rerank' || this.config.rag?.useReranking);
       if (!hasNumericPredicates && useReranking) {
         fullSources = (await this.reranker.rerank(fullSources, question, rerankLimit)).filter((s): s is VectorMatch => Boolean(s && typeof s === 'object'));
-      } else if (!wantsExhaustiveList) {
-        fullSources = fullSources.slice(0, topK);
       }
       const rerankMs = performance.now() - rerankStart;
 
-      // 4. Context Augmentation - Review all retrieved/reranked sources for full comprehension
-      let context = fullSources.length
-        ? fullSources.map((m, i) => `[Source ${i + 1}]\n${m?.content ?? ''}`).join('\n\n---\n\n')
+      // 4. Context Augmentation — Cap context sent to LLM to top 15 sources (~3000 tokens)
+      // to keep prompt under Groq's 6000 token limit while preserving ALL sources for UI citations.
+      const totalCount = fullSources.length;
+      const promptSources = fullSources.slice(0, 15);
+      let context = promptSources.length
+        ? `[SEARCH SUMMARY: Found ${totalCount} matching record(s) out of ${rawSources.length} total resources searched in database. Context includes top ${promptSources.length} relevant records]\n\n` +
+          promptSources.map((m, i) => `[Source ${i + 1}]\n${m?.content ?? ''}`).join('\n\n---\n\n')
         : 'No relevant context found.';
 
-      // Count the records of relevant sources to pass into the slice dynamically
-      let displayCount = 15; // default fallback
-
-      if (hasMetadataFilter) {
-        // If a metadata filter matches (e.g. brand, category), all of those matches are highly relevant
-        displayCount = fullSources.length;
-      } else {
-        // For general semantic queries, count the number of matches meeting high relevance threshold
-        const baseThreshold = Math.max(scoreThreshold, 0.4);
-        const highlyRelevant = fullSources.filter(m => (m?.score ?? 0) >= baseThreshold);
-        displayCount = Math.max(highlyRelevant.length, topK);
-        // Cap non-metadata semantic searches to a maximum of 15 to keep the UI clean
-        if (displayCount > 15) {
-          displayCount = 15;
-        }
-      }
-
-      // Ensure they are sorted from highest to lowest score, then slice using our dynamic count
+      // Do NOT truncate UI sources — keep all qualified matches for UI citations & tables
       const sources = [...fullSources]
         .filter((s): s is VectorMatch => Boolean(s && typeof s === 'object'))
-        .sort((a, b) => (b?.score ?? 0) - (a?.score ?? 0))
-        .slice(0, displayCount);
+        .sort((a, b) => (b?.score ?? 0) - (a?.score ?? 0));
 
       console.log('[Pipeline] Final sources for prompt & UI:', {
         count: sources.length,
-        sources: sources.map((s, idx) => ({
+        totalRawCount: rawSources.length,
+        sources: sources.slice(0, 10).map((s, idx) => ({
           rank: idx + 1,
           id: s?.id,
           score: s?.score,
@@ -864,10 +849,8 @@ You are a helpful assistant. Use the provided context to answer questions accura
         yield fullReply;
       }
 
-      // Kick off hallucination scoring now that we have the full reply.
-      // It runs in parallel with the remaining post-processing below.
-      const runHallucination = this.config.llm.options?.hallucinationScoring === true ||
-        (this.config.llm.provider !== 'ollama' && this.config.llm.options?.hallucinationScoring !== false);
+      // Kick off hallucination scoring ONLY when explicitly enabled by options
+      const runHallucination = this.config.llm.options?.hallucinationScoring === true;
       if (runHallucination) {
         hallucinationScoringPromise = scoreHallucination(this.llmProvider, fullReply, context)
           .catch(() => undefined);
@@ -1003,9 +986,8 @@ You are a helpful assistant. Use the provided context to answer questions accura
       );
     }
 
-    const isLocalProvider = this.config.llm.provider === 'ollama';
-    const disableLlmUiTransform = this.config.llm.options?.disableLlmUiTransform === true;
-    if (forceDeterministic || isLocalProvider || disableLlmUiTransform) {
+    const enableLlmUiTransform = this.config.llm.options?.enableLlmUiTransform === true;
+    if (forceDeterministic || !enableLlmUiTransform) {
       return UITransformer.transform(question, sources, this.config, cachedSchema);
     }
 
@@ -1211,7 +1193,7 @@ You are a helpful assistant. Use the provided context to answer questions accura
 Focus on extracting the core intent and entities. Do not answer the question, just rewrite it.
 
 History:
-${history.map(m => `${m.role}: ${m.content}`).join('\n')}
+${(history || []).map(m => `${m?.role || 'user'}: ${m?.content ?? ''}`).join('\n')}
 
 New Question: ${question}
 
@@ -1240,9 +1222,9 @@ Optimized Search Query:`;
 
     try {
       const { sources } = await this.retrieve(query, { namespace: ns, topK: 5 });
-      if (sources.length === 0) return [];
+      if (!sources || sources.length === 0) return [];
 
-      const context = sources.map((s) => s.content).join('\n\n---\n\n');
+      const context = sources.map((s) => s?.content ?? '').join('\n\n---\n\n');
       const prompt = `Based on the following snippets from a document, what are at least 5 short, relevant questions a user might ask? 
 Focus on questions that can be answered by the context.
 Keep each question under 10 words and make them very specific to the content.

package/src/llm/providers/GeminiProvider.ts

@@ -175,11 +175,11 @@ export class GeminiProvider implements ILLMProvider {
    *   - Messages must strictly alternate between `user` and `model`.
    */
   private buildGeminiContents(messages: ChatMessage[]): Array<{ role: 'user' | 'model'; parts: [{ text: string }] }> {
-    return messages
-      .filter((m) => m.role !== 'system')
+    return (messages || [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object' && m.role !== 'system'))
       .map((m) => ({
-        role: m.role === 'assistant' ? ('model' as const) : ('user' as const),
-        parts: [{ text: m.content }] as [{ text: string }],
+        role: m.role === 'assistant' ? 'model' : 'user',
+        parts: [{ text: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : '') }],
       }));
   }
 

package/src/llm/providers/GroqProvider.ts

@@ -95,12 +95,16 @@ export class GroqProvider implements ILLMProvider {
       content: buildSystemContent(basePrompt, context),
     };
 
+    const safeMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: (m.role || 'user') as 'user' | 'assistant',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
       systemMessage,
-      ...messages.map((m) => ({
-        role: m.role as 'user' | 'assistant',
-        content: m.content,
-      })),
+      ...safeMessages,
     ];
 
     const completion = await this.client.chat.completions.create({
@@ -125,12 +129,16 @@ export class GroqProvider implements ILLMProvider {
       content: buildSystemContent(basePrompt, context),
     };
 
+    const safeStreamMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: (m.role || 'user') as 'user' | 'assistant',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
       systemMessage,
-      ...messages.map((m) => ({
-        role: m.role as 'user' | 'assistant',
-        content: m.content,
-      })),
+      ...safeStreamMessages,
     ];
 
     const stream = await this.client.chat.completions.create({

package/src/llm/providers/OpenAIProvider.ts

@@ -91,12 +91,16 @@ export class OpenAIProvider implements ILLMProvider {
       content: buildSystemContent(basePrompt, context),
     };
 
+    const safeMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: (m.role || 'user') as 'user' | 'assistant',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
       systemMessage,
-      ...messages.map((m) => ({
-        role: m.role as 'user' | 'assistant',
-        content: m.content,
-      })),
+      ...safeMessages,
     ];
 
     const completion = await this.client.chat.completions.create({
@@ -121,12 +125,16 @@ export class OpenAIProvider implements ILLMProvider {
       content: buildSystemContent(basePrompt, context),
     };
 
+    const safeStreamMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: (m.role || 'user') as 'user' | 'assistant',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
       systemMessage,
-      ...messages.map((m) => ({
-        role: m.role as 'user' | 'assistant',
-        content: m.content,
-      })),
+      ...safeStreamMessages,
     ];
 
     const stream = await this.client.chat.completions.create({

package/src/llm/providers/QwenProvider.ts

@@ -99,12 +99,16 @@ export class QwenProvider implements ILLMProvider {
       content: buildSystemContent(basePrompt, context),
     };
 
+    const safeMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: (m.role || 'user') as 'user' | 'assistant',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
       systemMessage,
-      ...messages.map((m) => ({
-        role: m.role as 'user' | 'assistant',
-        content: m.content,
-      })),
+      ...safeMessages,
     ];
 
     const completion = await this.client.chat.completions.create({
@@ -129,12 +133,16 @@ export class QwenProvider implements ILLMProvider {
       content: buildSystemContent(basePrompt, context),
     };
 
+    const safeStreamMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: (m.role || 'user') as 'user' | 'assistant',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages: OpenAI.Chat.ChatCompletionMessageParam[] = [
       systemMessage,
-      ...messages.map((m) => ({
-        role: m.role as 'user' | 'assistant',
-        content: m.content,
-      })),
+      ...safeStreamMessages,
     ];
 
     const stream = await this.client.chat.completions.create({

package/src/llm/providers/UniversalLLMAdapter.ts

@@ -18,6 +18,17 @@ interface UniversalLLMOptions {
   timeout?: number;
 }
 
+export function extractContent(obj: unknown): string | undefined {
+  if (!obj || typeof obj !== 'object') return undefined;
+  const choice = (obj as any).choices?.[0];
+  if (!choice) return undefined;
+  if (typeof choice.message?.content === 'string') return choice.message.content;
+  if (typeof choice.delta?.content === 'string') return choice.delta.content;
+  if (typeof choice.text === 'string') return choice.text;
+  if (typeof choice.content === 'string') return choice.content;
+  return undefined;
+}
+
 export class UniversalLLMAdapter implements ILLMProvider {
   private readonly http: AxiosInstance;
   private readonly model: string;
@@ -72,9 +83,16 @@ export class UniversalLLMAdapter implements ILLMProvider {
   async chat(messages: ChatMessage[], context?: string): Promise<string> {
     const path = this.opts.chatPath ?? '/chat/completions';
 
+    const safeMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: m.role || 'user',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages = [
       { role: 'system', content: `${this.systemPrompt}\n\nContext:\n${context ?? 'None'}` },
-      ...messages,
+      ...safeMessages,
     ];
 
     let payload: unknown;
@@ -97,28 +115,44 @@ export class UniversalLLMAdapter implements ILLMProvider {
     const isSelfHost = this.baseUrl.includes('retrivora.com') || this.baseUrl.includes('localhost');
     if (isSelfHost || Boolean(process.env.VERCEL)) {
       const _g = globalThis as any;
-      if (typeof _g.__retrivoraDispatchChat === 'function') {
+      let dispatch = _g.__retrivoraDispatchChat;
+      if (typeof dispatch !== 'function') {
+        try {
+          const gateway = await import('../../../../../services/llm-gateway/router');
+          if (gateway?.dispatchChatCompletion) {
+            _g.__retrivoraDispatchChat = gateway.dispatchChatCompletion;
+            _g.__retrivoraDispatchEmbedding = gateway.dispatchEmbedding;
+            dispatch = gateway.dispatchChatCompletion;
+          }
+        } catch { /* proceed */ }
+      }
+
+      if (typeof dispatch === 'function') {
         try {
-          const res = await _g.__retrivoraDispatchChat({
+          const res = await dispatch({
             model: this.model,
             messages: formattedMessages as any,
             max_tokens: this.maxTokens,
             temperature: this.temperature,
           }, this.apiKey);
 
-          if (res?.response?.choices?.[0]?.message?.content) {
-            return res.response.choices[0].message.content;
+          const content = extractContent(res?.response);
+          if (content !== undefined) {
+            return content;
           }
+          throw new Error(`[UniversalLLMAdapter] In-process dispatch returned empty content for model: ${this.model}`);
         } catch (dispatchErr) {
           throw dispatchErr;
         }
+      } else {
+        throw new Error(`[UniversalLLMAdapter] In-process gateway dispatch not registered. Direct self-referential HTTP calls to ${this.baseUrl} are disabled on Vercel to prevent HTTP 508 Loop Detected.`);
       }
     }
 
     const { data } = await this.http.post(path, payload);
 
     const extractPath = this.opts.responseExtractPath ?? 'choices[0].message.content';
-    const result = resolvePath(data, extractPath);
+    const result = resolvePath(data, extractPath) ?? extractContent(data);
 
     if (result === undefined) {
       throw new Error(`[UniversalLLMAdapter] Could not extract text from path '${extractPath}' in response.`);
@@ -139,9 +173,16 @@ export class UniversalLLMAdapter implements ILLMProvider {
     const extractPath = (this.opts.responseExtractPath ?? 'choices[0].message.content')
       .replace('message.content', 'delta.content');
 
+    const safeMessages = (Array.isArray(messages) ? messages : [])
+      .filter((m): m is ChatMessage => Boolean(m && typeof m === 'object'))
+      .map((m) => ({
+        role: m.role || 'user',
+        content: typeof m.content === 'string' ? m.content : (m.content != null ? String(m.content) : ''),
+      }));
+
     const formattedMessages = [
       { role: 'system', content: `${this.systemPrompt}\n\nContext:\n${context ?? 'None'}` },
-      ...messages,
+      ...safeMessages,
     ];
 
     let payload: unknown;
@@ -171,9 +212,21 @@ export class UniversalLLMAdapter implements ILLMProvider {
 
     if (isSelfHost || Boolean(process.env.VERCEL)) {
       const _g = globalThis as any;
-      if (typeof _g.__retrivoraDispatchChat === 'function') {
+      let dispatch = _g.__retrivoraDispatchChat;
+      if (typeof dispatch !== 'function') {
         try {
-          const res = await _g.__retrivoraDispatchChat({
+          const gateway = await import('../../../../../services/llm-gateway/router');
+          if (gateway?.dispatchChatCompletion) {
+            _g.__retrivoraDispatchChat = gateway.dispatchChatCompletion;
+            _g.__retrivoraDispatchEmbedding = gateway.dispatchEmbedding;
+            dispatch = gateway.dispatchChatCompletion;
+          }
+        } catch { /* proceed */ }
+      }
+
+      if (typeof dispatch === 'function') {
+        try {
+          const res = await dispatch({
             model: this.model,
             messages: formattedMessages as any,
             max_tokens: this.maxTokens,
@@ -183,15 +236,19 @@ export class UniversalLLMAdapter implements ILLMProvider {
 
           if (res?.stream) {
             streamBody = res.stream as ReadableStream<Uint8Array>;
-          } else if (res?.response?.choices?.[0]?.message?.content) {
-            yield res.response.choices[0].message.content;
-            return;
+          } else {
+            const content = extractContent(res?.response);
+            if (content !== undefined) {
+              yield content;
+              return;
+            }
+            throw new Error(`[UniversalLLMAdapter] In-process dispatch stream returned empty response for model: ${this.model}`);
           }
         } catch (dispatchErr) {
-          // If in-process dispatch threw a real error (Rate limit, Auth, API error), rethrow it
-          // rather than silently falling back to external HTTP fetch which masks the error.
           throw dispatchErr;
         }
+      } else {
+        throw new Error(`[UniversalLLMAdapter] In-process gateway dispatch not registered. Direct self-referential HTTP calls to ${this.baseUrl} are disabled on Vercel to prevent HTTP 508 Loop Detected.`);
       }
     }
 
@@ -233,7 +290,7 @@ export class UniversalLLMAdapter implements ILLMProvider {
 
           try {
             const json = JSON.parse(trimmed.slice(5).trim());
-            const text = resolvePath(json, extractPath);
+            const text = (resolvePath(json, extractPath) as string | undefined) ?? extractContent(json);
             if (text && typeof text === 'string') yield text;
           } catch {
             // Malformed SSE line — skip
@@ -246,7 +303,7 @@ export class UniversalLLMAdapter implements ILLMProvider {
         const jsonStr = buffer.replace(/^data:\s*/, '').trim();
         try {
           const json = JSON.parse(jsonStr);
-          const text = resolvePath(json, extractPath);
+          const text = (resolvePath(json, extractPath) as string | undefined) ?? extractContent(json);
           if (text && typeof text === 'string') yield text;
         } catch { /* ignore final partial line */ }
       }
@@ -274,9 +331,21 @@ export class UniversalLLMAdapter implements ILLMProvider {
     const isSelfHost = this.baseUrl.includes('retrivora.com') || this.baseUrl.includes('localhost');
     if (isSelfHost || Boolean(process.env.VERCEL)) {
       const _g = globalThis as any;
-      if (typeof _g.__retrivoraDispatchEmbedding === 'function') {
+      let dispatch = _g.__retrivoraDispatchEmbedding;
+      if (typeof dispatch !== 'function') {
+        try {
+          const gateway = await import('../../../../../services/llm-gateway/router');
+          if (gateway?.dispatchEmbedding) {
+            _g.__retrivoraDispatchChat = gateway.dispatchChatCompletion;
+            _g.__retrivoraDispatchEmbedding = gateway.dispatchEmbedding;
+            dispatch = gateway.dispatchEmbedding;
+          }
+        } catch { /* proceed */ }
+      }
+
+      if (typeof dispatch === 'function') {
         try {
-          const res = await _g.__retrivoraDispatchEmbedding({
+          const res = await dispatch({
             model: this.model,
             input: text,
           }, this.apiKey);
@@ -284,9 +353,12 @@ export class UniversalLLMAdapter implements ILLMProvider {
           if (res?.data?.[0]?.embedding) {
             return res.data[0].embedding;
           }
+          throw new Error(`[UniversalLLMAdapter] In-process embedding dispatch returned invalid vector for model: ${this.model}`);
         } catch (dispatchErr) {
           throw dispatchErr;
         }
+      } else {
+        throw new Error(`[UniversalLLMAdapter] In-process gateway dispatch not registered. Direct self-referential HTTP calls to ${this.baseUrl} are disabled on Vercel to prevent HTTP 508 Loop Detected.`);
       }
     }
 

package/src/rendering/IntentClassifier.ts

@@ -126,10 +126,10 @@ export class IntentClassifier {
   // ─── Intent Heuristic Checkers ─────────────────────────────────────────────
 
   private static isInformationLookup(query: string): boolean {
-    // Exact single-fact pattern matches e.g. "What is X?", "Explain Y", "Price of Z", "Can Auvira be assigned..."
-    if (/^(what|who|explain|define|where|when|why|how|how does|can|could|does|do|is|are|has|have|will|should|would|meaning of)\s+/i.test(query)) {
+    // Exact single-fact pattern matches e.g. "What is X?", "Explain Y", "Explore Z", "Tell me about A"
+    if (/^(what|who|explain|explore|define|where|when|why|how|how does|can|could|does|do|is|are|has|have|will|should|would|meaning of|tell me about|show details of|overview of|summarize)\s+/i.test(query)) {
       // Exclude comparison or analytical phrasing inside "what is the trend..."
-      if (!/\b(compare|versus|vs\.?|trend|monthly|revenue|top selling|best)\b/i.test(query)) {
+      if (!/\b(compare|versus|vs\.?|trend|monthly|revenue|top 10|top selling|best|chart|graph|plot|breakdown|analytics)\b/i.test(query)) {
         return true;
       }
     }

package/src/rendering/rules/Rule1SpecificInfoRule.ts

@@ -10,7 +10,7 @@ export class Rule1SpecificInfoRule implements IRenderRule {
     // Check specific fact/definition/explanation indicators
     const isFactQuery =
       intent === 'information_lookup' ||
-      /^(what|who|explain|define|where|when|why|meaning of)\s+/i.test(q) ||
+      /^(what|who|explain|explore|define|where|when|why|meaning of|tell me about|show details of|overview of|summarize)\s+/i.test(q) ||
       /^(what is the price of|price of|cost of)\b/i.test(q);
 
     const isNonVisual = !/\b(compare|versus|vs\.?|trend|monthly|revenue|best|top 10|chart|table|list|grid)\b/i.test(q);

package/src/utils/ProductExtractor.ts

@@ -196,6 +196,7 @@ export function isLikelyUiOnlyMessage(text: string): boolean {
 export function extractProductsFromSources(sources: VectorMatch[] | undefined, isUser: boolean): Product[] {
   if (isUser || !sources) return [];
   return sources
+    .filter((s): s is VectorMatch => Boolean(s && typeof s === 'object'))
     .filter(s => {
       const m = s.metadata ?? {};
       const keys = Object.keys(m).map(k => k.toLowerCase());
@@ -204,7 +205,7 @@ export function extractProductsFromSources(sources: VectorMatch[] | undefined, i
       );
       const hasProductIdentity = keys.some(k => ['brand', 'model'].includes(k)) &&
         keys.some(k => ['name', 'title', 'product_name', 'product'].includes(k));
-      const hasPricePattern = /\$\s*\d+/.test(s.content);
+      const hasPricePattern = /\$\s*\d+/.test(s?.content ?? '');
       return hasStrongProductKey || hasProductIdentity || hasPricePattern || m.type === 'product';
     })
     .map(s => {
@@ -215,13 +216,13 @@ export function extractProductsFromSources(sources: VectorMatch[] | undefined, i
       const description = resolveMetadataValue(m, 'description') as string;
 
       return {
-        id: s.id,
+        id: s?.id ?? `prod_${Math.random().toString(36).substring(2, 8)}`,
         name: name ?? (s?.content ?? '').split('\n')[0] ?? 'Unknown Product',
         brand: brand,
         price: price,
         image: resolveImage(m),
         link: resolveMetadataValue(m, 'link') as string,
-        description: description ? description : '' //description ?? s.content,
+        description: description ? description : '',
       };
     });
 }