@retrivora-ai/rag-engine 2.0.7 → 2.0.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "2.0.7",
+  "version": "2.0.8",
   "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/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);
     }
 

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;
@@ -97,28 +108,44 @@ 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 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,
             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.`);
@@ -171,9 +198,21 @@ 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 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,
@@ -183,15 +222,19 @@ 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 +276,7 @@ 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 +289,7 @@ 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 +317,21 @@ 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 +339,12 @@ 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/utils/UITransformer.ts

@@ -560,8 +560,7 @@ RULES:
   ): UITransformationResponse {
     const products: CarouselProduct[] = data
       .map(item => this.extractProductInfo(item, config, trainedSchema))
-      .filter((p): p is CarouselProduct => p !== null)
-      .slice(0, 15);
+      .filter((p): p is CarouselProduct => p !== null);
 
     return {
       type: 'product_carousel',
@@ -1035,7 +1034,7 @@ RULES:
   }
 
   private static profileData(data: VectorMatch[]): DataProfile {
-    const records = data.map((item): DataRecord => {
+    const records = (data || []).filter((item): item is VectorMatch => Boolean(item && typeof item === 'object')).map((item): DataRecord => {
       const fields: Record<string, PrimitiveValue> = {};
       Object.entries(item.metadata || {}).forEach(([key, value]) => {
         const primitive = this.toPrimitive(value);
@@ -1044,8 +1043,8 @@ RULES:
       if (!fields.content && item.content) fields.content = item.content.substring(0, 500);
       return {
         id: item.id,
-        content: item.content,
-        score: item.score,
+        content: item.content ?? '',
+        score: item.score ?? 0,
         fields,
         source: item,
       };