@retrivora-ai/rag-engine 1.0.2 → 1.0.4

package/dist/{chunk-OJNAKSQ2.mjs → chunk-6MLZHQZT.mjs}

@@ -1391,29 +1391,59 @@ var EntityExtractor = class {
   async extract(text) {
     const prompt = `
 Extract entities and relationships from the following text.
-Format the output as JSON with two keys: "nodes" (array of {id, label, properties}) and "edges" (array of {source, target, type, properties}).
-Use the same ID for the same entity.
+Format the output as a JSON object with exactly two keys: "nodes" and "edges".
 
-IMPORTANT: Ensure all property values are valid JSON types (strings, numbers, booleans, or null). 
-DO NOT include mathematical expressions like "4.5/5" as numbers; use strings instead.
+Nodes: {id, label, properties}
+Edges: {source, target, type, properties}
 
-Text:
-"${text}"
+IMPORTANT:
+- Ensure all property values are simple JSON types (string, number, boolean).
+- Do not use mathematical expressions or complex nested objects in properties.
+- If no entities are found, return empty arrays.
+- RESPOND ONLY WITH THE JSON OBJECT.
 
-Output JSON:
+Text to extract from:
+"${text}"
 `;
     const response = await this.llm.chat([
-      { role: "system", content: "You are an expert at knowledge graph extraction. Respond only with valid JSON." },
+      { role: "system", content: "You are a precise knowledge graph extraction engine. You always output valid JSON and nothing else." },
       { role: "user", content: prompt }
-    ], "");
+    ], "", { maxTokens: 4096, temperature: 0.1 });
     try {
-      const cleanJson = response.replace(/```json\n?|\n?```/g, "").trim();
+      const jsonMatch = response.match(/\{[\s\S]*\}/);
+      const cleanJson = jsonMatch ? jsonMatch[0] : response.trim();
       return JSON.parse(cleanJson);
     } catch (e) {
-      console.warn("[EntityExtractor] Failed to parse LLM response as JSON:", response);
+      try {
+        const repaired = this.repairTruncatedJson(response);
+        if (repaired) return JSON.parse(repaired);
+      } catch (e2) {
+      }
+      console.warn(`[EntityExtractor] Failed to parse LLM response. Length: ${response.length} chars.`);
+      console.warn("[EntityExtractor] Snippet:", response.substring(0, 100) + "...");
       return { nodes: [], edges: [] };
     }
   }
+  /**
+   * Attempts to fix JSON that was cut off mid-generation.
+   * Strategy: Find the last valid element in an array and close the structure.
+   */
+  repairTruncatedJson(json) {
+    let text = json.trim();
+    const startIdx = text.indexOf("{");
+    if (startIdx === -1) return null;
+    text = text.substring(startIdx);
+    const lastCompleteObjectIdx = text.lastIndexOf("}");
+    if (lastCompleteObjectIdx === -1) return null;
+    let repaired = text.substring(0, lastCompleteObjectIdx + 1);
+    const openBrackets = (repaired.match(/\{/g) || []).length;
+    const closedBrackets = (repaired.match(/\}/g) || []).length;
+    const openSquares = (repaired.match(/\[/g) || []).length;
+    const closedSquares = (repaired.match(/\]/g) || []).length;
+    for (let i = 0; i < openSquares - closedSquares; i++) repaired += "]";
+    for (let i = 0; i < openBrackets - closedBrackets; i++) repaired += "}";
+    return repaired;
+  }
 };
 
 // src/rag/Reranker.ts

package/dist/handlers/index.js

@@ -2854,29 +2854,59 @@ var EntityExtractor = class {
   async extract(text) {
     const prompt = `
 Extract entities and relationships from the following text.
-Format the output as JSON with two keys: "nodes" (array of {id, label, properties}) and "edges" (array of {source, target, type, properties}).
-Use the same ID for the same entity.
+Format the output as a JSON object with exactly two keys: "nodes" and "edges".
 
-IMPORTANT: Ensure all property values are valid JSON types (strings, numbers, booleans, or null). 
-DO NOT include mathematical expressions like "4.5/5" as numbers; use strings instead.
+Nodes: {id, label, properties}
+Edges: {source, target, type, properties}
 
-Text:
-"${text}"
+IMPORTANT:
+- Ensure all property values are simple JSON types (string, number, boolean).
+- Do not use mathematical expressions or complex nested objects in properties.
+- If no entities are found, return empty arrays.
+- RESPOND ONLY WITH THE JSON OBJECT.
 
-Output JSON:
+Text to extract from:
+"${text}"
 `;
     const response = await this.llm.chat([
-      { role: "system", content: "You are an expert at knowledge graph extraction. Respond only with valid JSON." },
+      { role: "system", content: "You are a precise knowledge graph extraction engine. You always output valid JSON and nothing else." },
       { role: "user", content: prompt }
-    ], "");
+    ], "", { maxTokens: 4096, temperature: 0.1 });
     try {
-      const cleanJson = response.replace(/```json\n?|\n?```/g, "").trim();
+      const jsonMatch = response.match(/\{[\s\S]*\}/);
+      const cleanJson = jsonMatch ? jsonMatch[0] : response.trim();
       return JSON.parse(cleanJson);
     } catch (e) {
-      console.warn("[EntityExtractor] Failed to parse LLM response as JSON:", response);
+      try {
+        const repaired = this.repairTruncatedJson(response);
+        if (repaired) return JSON.parse(repaired);
+      } catch (e2) {
+      }
+      console.warn(`[EntityExtractor] Failed to parse LLM response. Length: ${response.length} chars.`);
+      console.warn("[EntityExtractor] Snippet:", response.substring(0, 100) + "...");
       return { nodes: [], edges: [] };
     }
   }
+  /**
+   * Attempts to fix JSON that was cut off mid-generation.
+   * Strategy: Find the last valid element in an array and close the structure.
+   */
+  repairTruncatedJson(json) {
+    let text = json.trim();
+    const startIdx = text.indexOf("{");
+    if (startIdx === -1) return null;
+    text = text.substring(startIdx);
+    const lastCompleteObjectIdx = text.lastIndexOf("}");
+    if (lastCompleteObjectIdx === -1) return null;
+    let repaired = text.substring(0, lastCompleteObjectIdx + 1);
+    const openBrackets = (repaired.match(/\{/g) || []).length;
+    const closedBrackets = (repaired.match(/\}/g) || []).length;
+    const openSquares = (repaired.match(/\[/g) || []).length;
+    const closedSquares = (repaired.match(/\]/g) || []).length;
+    for (let i = 0; i < openSquares - closedSquares; i++) repaired += "]";
+    for (let i = 0; i < openBrackets - closedBrackets; i++) repaired += "}";
+    return repaired;
+  }
 };
 
 // src/rag/Reranker.ts

package/dist/handlers/index.mjs

@@ -4,7 +4,7 @@ import {
   createIngestHandler,
   createStreamHandler,
   createUploadHandler
-} from "../chunk-OJNAKSQ2.mjs";
+} from "../chunk-6MLZHQZT.mjs";
 import "../chunk-YLTMFW4M.mjs";
 import "../chunk-X4TOT24V.mjs";
 export {

package/dist/server.js

@@ -2936,29 +2936,59 @@ var EntityExtractor = class {
   async extract(text) {
     const prompt = `
 Extract entities and relationships from the following text.
-Format the output as JSON with two keys: "nodes" (array of {id, label, properties}) and "edges" (array of {source, target, type, properties}).
-Use the same ID for the same entity.
+Format the output as a JSON object with exactly two keys: "nodes" and "edges".
 
-IMPORTANT: Ensure all property values are valid JSON types (strings, numbers, booleans, or null). 
-DO NOT include mathematical expressions like "4.5/5" as numbers; use strings instead.
+Nodes: {id, label, properties}
+Edges: {source, target, type, properties}
 
-Text:
-"${text}"
+IMPORTANT:
+- Ensure all property values are simple JSON types (string, number, boolean).
+- Do not use mathematical expressions or complex nested objects in properties.
+- If no entities are found, return empty arrays.
+- RESPOND ONLY WITH THE JSON OBJECT.
 
-Output JSON:
+Text to extract from:
+"${text}"
 `;
     const response = await this.llm.chat([
-      { role: "system", content: "You are an expert at knowledge graph extraction. Respond only with valid JSON." },
+      { role: "system", content: "You are a precise knowledge graph extraction engine. You always output valid JSON and nothing else." },
       { role: "user", content: prompt }
-    ], "");
+    ], "", { maxTokens: 4096, temperature: 0.1 });
     try {
-      const cleanJson = response.replace(/```json\n?|\n?```/g, "").trim();
+      const jsonMatch = response.match(/\{[\s\S]*\}/);
+      const cleanJson = jsonMatch ? jsonMatch[0] : response.trim();
       return JSON.parse(cleanJson);
     } catch (e) {
-      console.warn("[EntityExtractor] Failed to parse LLM response as JSON:", response);
+      try {
+        const repaired = this.repairTruncatedJson(response);
+        if (repaired) return JSON.parse(repaired);
+      } catch (e2) {
+      }
+      console.warn(`[EntityExtractor] Failed to parse LLM response. Length: ${response.length} chars.`);
+      console.warn("[EntityExtractor] Snippet:", response.substring(0, 100) + "...");
       return { nodes: [], edges: [] };
     }
   }
+  /**
+   * Attempts to fix JSON that was cut off mid-generation.
+   * Strategy: Find the last valid element in an array and close the structure.
+   */
+  repairTruncatedJson(json) {
+    let text = json.trim();
+    const startIdx = text.indexOf("{");
+    if (startIdx === -1) return null;
+    text = text.substring(startIdx);
+    const lastCompleteObjectIdx = text.lastIndexOf("}");
+    if (lastCompleteObjectIdx === -1) return null;
+    let repaired = text.substring(0, lastCompleteObjectIdx + 1);
+    const openBrackets = (repaired.match(/\{/g) || []).length;
+    const closedBrackets = (repaired.match(/\}/g) || []).length;
+    const openSquares = (repaired.match(/\[/g) || []).length;
+    const closedSquares = (repaired.match(/\]/g) || []).length;
+    for (let i = 0; i < openSquares - closedSquares; i++) repaired += "]";
+    for (let i = 0; i < openBrackets - closedBrackets; i++) repaired += "}";
+    return repaired;
+  }
 };
 
 // src/rag/Reranker.ts

package/dist/server.mjs

@@ -34,7 +34,7 @@ import {
   createIngestHandler,
   createUploadHandler,
   getRagConfig
-} from "./chunk-OJNAKSQ2.mjs";
+} from "./chunk-6MLZHQZT.mjs";
 import "./chunk-YLTMFW4M.mjs";
 import {
   PineconeProvider

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@retrivora-ai/rag-engine",
-  "version": "1.0.2",
+  "version": "1.0.4",
   "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": "MIT",

package/src/rag/EntityExtractor.ts

@@ -17,30 +17,74 @@ export class EntityExtractor {
   async extract(text: string): Promise<{ nodes: GraphNode[]; edges: Edge[] }> {
     const prompt = `
 Extract entities and relationships from the following text.
-Format the output as JSON with two keys: "nodes" (array of {id, label, properties}) and "edges" (array of {source, target, type, properties}).
-Use the same ID for the same entity.
+Format the output as a JSON object with exactly two keys: "nodes" and "edges".
 
-IMPORTANT: Ensure all property values are valid JSON types (strings, numbers, booleans, or null). 
-DO NOT include mathematical expressions like "4.5/5" as numbers; use strings instead.
+Nodes: {id, label, properties}
+Edges: {source, target, type, properties}
 
-Text:
-"${text}"
+IMPORTANT:
+- Ensure all property values are simple JSON types (string, number, boolean).
+- Do not use mathematical expressions or complex nested objects in properties.
+- If no entities are found, return empty arrays.
+- RESPOND ONLY WITH THE JSON OBJECT.
 
-Output JSON:
+Text to extract from:
+"${text}"
 `;
 
     const response = await this.llm.chat([
-      { role: 'system', content: 'You are an expert at knowledge graph extraction. Respond only with valid JSON.' },
+      { role: 'system', content: 'You are a precise knowledge graph extraction engine. You always output valid JSON and nothing else.' },
       { role: 'user', content: prompt }
-    ], '');
+    ], '', { maxTokens: 4096, temperature: 0.1 });
 
     try {
-      // Clean up LLM response (sometimes it adds markdown blocks)
-      const cleanJson = response.replace(/```json\n?|\n?```/g, '').trim();
+      // 1. Try to find the JSON block using a regex
+      const jsonMatch = response.match(/\{[\s\S]*\}/);
+      const cleanJson = jsonMatch ? jsonMatch[0] : response.trim();
+      
       return JSON.parse(cleanJson);
     } catch {
-      console.warn('[EntityExtractor] Failed to parse LLM response as JSON:', response);
+      // 2. If parsing fails, try to repair truncated JSON
+      try {
+        const repaired = this.repairTruncatedJson(response);
+        if (repaired) return JSON.parse(repaired);
+      } catch { /* ignore repair failure */ }
+
+      console.warn(`[EntityExtractor] Failed to parse LLM response. Length: ${response.length} chars.`);
+      console.warn('[EntityExtractor] Snippet:', response.substring(0, 100) + '...');
       return { nodes: [], edges: [] };
     }
   }
+
+  /**
+   * Attempts to fix JSON that was cut off mid-generation.
+   * Strategy: Find the last valid element in an array and close the structure.
+   */
+  private repairTruncatedJson(json: string): string | null {
+    let text = json.trim();
+    
+    // Find the first { and work from there
+    const startIdx = text.indexOf('{');
+    if (startIdx === -1) return null;
+    text = text.substring(startIdx);
+
+    // If it's truncated in an array like "nodes": [ ... , { ...
+    // We try to find the last complete object before the truncation
+    const lastCompleteObjectIdx = text.lastIndexOf('}');
+    if (lastCompleteObjectIdx === -1) return null;
+
+    // Truncate at the last complete object
+    let repaired = text.substring(0, lastCompleteObjectIdx + 1);
+
+    // Close any open structures (arrays or the main object)
+    const openBrackets = (repaired.match(/\{/g) || []).length;
+    const closedBrackets = (repaired.match(/\}/g) || []).length;
+    const openSquares = (repaired.match(/\[/g) || []).length;
+    const closedSquares = (repaired.match(/\]/g) || []).length;
+
+    for (let i = 0; i < openSquares - closedSquares; i++) repaired += ']';
+    for (let i = 0; i < openBrackets - closedBrackets; i++) repaired += '}';
+
+    return repaired;
+  }
 }