@tiledesk/tiledesk-server 2.16.0 → 2.17.3

package/CHANGELOG.md

@@ -5,6 +5,25 @@
 🚀        IN PRODUCTION                        🚀
 (https://www.npmjs.com/package/@tiledesk/tiledesk-server/v/2.3.77) 
 
+# 2.17.3
+- Added missing import path on kb route
+
+# 2.17.2
+- Added support for situated context in kb route
+- Added RAG context management to KB routes
+- Added support for scrape type 0 (alias: trafilatura)
+
+# 2.16.2
+- Improved multiplier retrieval for model types in quotes route
+
+# 2.16.1
+- Added stream option support to the KB /qa endpoint for real-time responses
+- Enhanced file upload route to correctly handle .webm files
+- Optimized token consumption and management in knowledge base operations
+
+# 2.16.0-hf
+- Fixed bug: issue on audio sent from widget
+
 # 2.16.0
 - Added possibility to update Knowledge Base content
 - Added rated only filter in Conversations History

package/config/kb/prompt/rag/PromptManager.js

@@ -0,0 +1,57 @@
+const fs = require('fs');
+const path = require('path');
+
+const modelMap = {
+    "gpt-3.5-turbo":        "gpt-3.5.txt",
+    "gpt-4":                "gpt-4.txt",
+    "gpt-4-turbo-preview":  "gpt-4.txt",
+    "gpt-4o":               "gpt-4o.txt",
+    "gpt-4o-mini":          "gpt-4o.txt",
+    "gpt-4.1":              "gpt-4.1.txt",
+    "gpt-4.1-mini":         "gpt-4.1.txt",
+    "gpt-4.1-nano":         "gpt-4.1.txt",
+    "gpt-5":                "gpt-5.txt",
+    "gpt-5-mini":           "gpt-5.txt",
+    "gpt-5-nano":           "gpt-5.txt",
+    "gpt-5.1":              "gpt-5.x.txt",
+    "gpt-5.2":              "gpt-5.x.txt",
+    "gpt-5.3-chat-latest":  "gpt-5.x.txt",
+    "gpt-5.4":              "gpt-5.x.txt",
+    "gpt-5.4-mini":         "gpt-5.x.txt",
+    "gpt-5.4-nano":         "gpt-5.x.txt",
+    "general":              "general.txt"
+}
+
+
+class PromptManager {
+  
+    constructor(basePath) {
+    this.basePath = basePath;
+    this.cache = new Map();
+  }
+
+  getPrompt(name) {
+    if (this.cache.has(name)) {
+      return this.cache.get(name);
+    }
+
+    const fileName = modelMap[name] || modelMap["general"];
+    const filePath = path.join(this.basePath, fileName);
+
+    let content;
+    try {
+      content = fs.readFileSync(filePath, 'utf-8');
+    } catch (err) {
+      content = fs.readFileSync(
+        path.join(this.basePath, modelMap["general"]),
+        'utf-8'
+      );
+    }
+
+    this.cache.set(name, content);
+    return content;
+  }
+}
+
+PromptManager.modelMap = modelMap;
+module.exports = PromptManager;

package/config/kb/prompt/rag/general.txt

@@ -0,0 +1,9 @@
+You are an helpful assistant for question-answering tasks. Follow these steps carefully:
+
+1. Answer in the same language of the user question, regardless of the retrieved context language
+2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.
+3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer.
+
+==Retrieved context start==
+{context}
+==Retrieved context end==

package/config/kb/prompt/rag/gpt-3.5.txt

@@ -0,0 +1,9 @@
+You are an helpful assistant for question-answering tasks.
+
+Use ONLY the pieces of retrieved context delimited by #### and the chat history to answer the question.
+
+If you don't know the answer, just say: "I don't know<NOANS>"
+
+####
+{context}
+####

package/config/kb/prompt/rag/gpt-4.1.txt

@@ -0,0 +1,9 @@
+You are an helpful assistant for question-answering tasks. Follow these steps carefully:
+
+1. Answer in the same language of the user question, regardless of the retrieved context language
+2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.
+3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer
+
+==Retrieved context start==
+{context}
+==Retrieved context end==

package/config/kb/prompt/rag/gpt-4.txt

@@ -0,0 +1,11 @@
+You are an helpful assistant for question-answering tasks.
+
+Use ONLY the pieces of retrieved context delimited by #### and the chat history to answer the question.
+
+If you don't know the answer, just say that you don't know.
+
+If and only if none of the retrieved context is useful for your task, add this word to the end <NOANS>
+
+####
+{context}
+####

package/config/kb/prompt/rag/gpt-4o.txt

@@ -0,0 +1,9 @@
+You are an helpful assistant for question-answering tasks. Follow these steps carefully:
+
+1. Answer in the same language of the user question, regardless of the retrieved context language
+2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.
+3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, return <NOANS>.
+
+==Retrieved context start==
+{context}
+==Retrieved context end==

package/config/kb/prompt/rag/gpt-5.txt

@@ -0,0 +1,32 @@
+# ROLE
+You are an AI assistant that answers the user's question using only the information contained in the provided context.
+
+# LANGUAGE
+Answer in the same language as the user's question.
+
+# CONTEXT
+You will receive a context delimited by ######:
+######
+{context}
+######
+
+# INSTRUCTIONS
+- Use only the information explicitly contained in the context.
+- Answer the user's question directly, as a human assistant would.
+- Do not mention the context, the document, the source, or the fact that information was provided.
+- Do not say phrases such as:
+  - "according to the context"
+  - "in the provided context"
+  - "the document says"
+  - "based on the information provided"
+- Do not explain your reasoning.
+- Do not repeat the question.
+- Keep the answer concise, clear, and natural.
+- Do not add assumptions, external knowledge, or details not supported by the context.
+
+# FALLBACK
+If the context does not contain enough information to answer the question, reply with exactly:
+<NOANS>
+
+# OUTPUT
+Return only the final answer, with no preamble and no meta-commentary.

package/config/kb/prompt/rag/gpt-5.x.txt

@@ -0,0 +1,32 @@
+# ROLE
+You are an AI assistant that answers the user's question using only the information contained in the provided context.
+
+# LANGUAGE
+Answer in the same language as the user's question.
+
+# CONTEXT
+You will receive a context delimited by ######:
+######
+{context}
+######
+
+# INSTRUCTIONS
+- Use only the information explicitly contained in the context.
+- Answer the user's question directly, as a human assistant would.
+- Do not mention the context, the document, the source, or the fact that information was provided.
+- Do not say phrases such as:
+  - "according to the context"
+  - "in the provided context"
+  - "the document says"
+  - "based on the information provided"
+- Do not explain your reasoning.
+- Do not repeat the question.
+- Keep the answer concise, clear, and natural.
+- Do not add assumptions, external knowledge, or details not supported by the context.
+
+# FALLBACK
+If the context does not contain enough information to answer the question, reply with exactly:
+<NOANS>
+
+# OUTPUT
+Return only the final answer, with no preamble and no meta-commentary.

package/config/kb/situatedContext.js

@@ -0,0 +1,6 @@
+module.exports = {
+    enable: process.env.SITUATED_CONTEXT_ENABLE === "true",
+    provider: process.env.SITUATED_CONTEXT_PROVIDER || "openai",
+    model: process.env.SITUATED_CONTEXT_MODEL || "gpt-4o",
+    api_key: ""
+}

package/middleware/file-type.js

@@ -31,9 +31,10 @@ function areMimeTypesEquivalent(mimeType1, mimeType2) {
     'audio/wave': ['audio/wav', 'audio/x-wav', 'audio/vnd.wave'],
     'audio/x-wav': ['audio/wav', 'audio/wave', 'audio/vnd.wave'],
     'audio/vnd.wave': ['audio/wav', 'audio/wave', 'audio/x-wav'],
-    'audio/mpeg': ['audio/opus', 'audio/mp3'],
-    'audio/mp3': ['audio/mpeg', 'audio/opus'],
-    'audio/opus': ['audio/mpeg', 'audio/mp3'],
+    'audio/mpeg': ['audio/opus', 'audio/mp3', 'audio/webm'],
+    'audio/mp3': ['audio/mpeg', 'audio/opus', 'audio/webm'],
+    'audio/opus': ['audio/mpeg', 'audio/mp3', 'audio/webm'],
+    'audio/webm': ['audio/mpeg', 'audio/mp3', 'audio/opus'],
     'image/jpeg': ['image/jpg'],
     'image/jpg': ['image/jpeg'],
     'application/x-zip-compressed': ['application/zip'],
@@ -47,54 +48,126 @@ function areMimeTypesEquivalent(mimeType1, mimeType2) {
   return false;
 }
 
+// Magic bytes for fallback when file-type throws (e.g. strtok3/token-types Uint8Array vs Buffer)
+const MAGIC_SIGNATURES = {
+  'video/webm': [[0x1A, 0x45, 0xDF, 0xA3]],           // EBML
+  'audio/webm': [[0x1A, 0x45, 0xDF, 0xA3]],
+  'audio/mpeg': [[0xFF, 0xFB], [0xFF, 0xFA], [0xFF, 0xF3], [0xFF, 0xF2], [0x49, 0x44, 0x33]], // ID3 or MP3 frame
+  'audio/mp3': [[0xFF, 0xFB], [0xFF, 0xFA], [0xFF, 0xF3], [0xFF, 0xF2], [0x49, 0x44, 0x33]],
+  'image/png': [[0x89, 0x50, 0x4E, 0x47, 0x0D, 0x0A, 0x1A, 0x0A]],
+  'image/jpeg': [[0xFF, 0xD8, 0xFF]],
+  'image/gif': [[0x47, 0x49, 0x46, 0x38, 0x37, 0x61], [0x47, 0x49, 0x46, 0x38, 0x39, 0x61]],
+  'application/pdf': [[0x25, 0x50, 0x44, 0x46]],
+};
+
+function magicMatches(buf, mimetype) {
+  const signatures = MAGIC_SIGNATURES[mimetype && mimetype.toLowerCase()];
+  if (!signatures) return false;
+  for (const sig of signatures) {
+    if (buf.length < sig.length) continue;
+    let ok = true;
+    for (let i = 0; i < sig.length; i++) {
+      const b = buf[i] !== undefined ? (buf[i] & 0xFF) : -1;
+      if (b !== sig[i]) { ok = false; break; }
+    }
+    if (ok) return true;
+  }
+  return false;
+}
+
+const BASE64_REGEX = /^[A-Za-z0-9+/]+=*$/;
+
+/**
+ * Ensures the input is a Node.js Buffer. file-type (and token-types/strtok3) require
+ * a Buffer with methods like readUInt8; GridFS or other sources may return
+ * Uint8Array, ArrayBuffer, BSON Binary, or (when client sends base64) a string.
+ * We always allocate a new Buffer and copy bytes so file-type never receives
+ * a buffer-like that loses readUInt8 when sliced (e.g. by strtok3).
+ */
+function ensureBuffer(buffer) {
+  if (!buffer) return buffer;
+
+  // Base64 string (e.g. client sends form body as base64): decode to binary
+  if (typeof buffer === 'string' && buffer.length > 0) {
+    const trimmed = buffer.replace(/\s/g, '');
+    if (BASE64_REGEX.test(trimmed)) {
+      return Buffer.from(trimmed, 'base64');
+    }
+    return Buffer.from(buffer, 'utf8');
+  }
+
+  // Copy into a new Buffer so file-type's internal slices are always real Buffers
+  let uint8;
+  if (buffer instanceof Uint8Array) {
+    uint8 = buffer;
+  } else if (buffer instanceof ArrayBuffer) {
+    uint8 = new Uint8Array(buffer);
+  } else if (buffer && typeof buffer.buffer === 'object' && buffer.buffer instanceof ArrayBuffer) {
+    uint8 = new Uint8Array(buffer.buffer, buffer.byteOffset, buffer.byteLength);
+  } else if (Buffer.isBuffer(buffer)) {
+    uint8 = new Uint8Array(buffer.buffer, buffer.byteOffset, buffer.byteLength);
+  } else {
+    uint8 = new Uint8Array(Buffer.from(buffer));
+  }
+  return Buffer.from(uint8);
+}
+
 async function verifyFileContent(buffer, mimetype) {
   if (!buffer) throw new Error("No file provided");
 
+  const buf = ensureBuffer(buffer);
+
+  let fileType;
   try {
-    const fileType = await FileType.fromBuffer(buffer);
+    fileType = await FileType.fromBuffer(buf);
+  } catch (err) {
+    // strtok3 uses Uint8Array for numBuffer but token-types expects Buffer.readUInt8 (known compat bug in deps)
+    if (err && typeof err.message === 'string' && err.message.includes('readUInt8')) {
+      if (mimetype && magicMatches(buf, mimetype)) return true;
+      const err2 = new Error(`File content could not be verified. Declared mimetype: ${mimetype}`);
+      err2.source = "FileContentVerification";
+      throw err2;
+    }
+    throw err;
+  }
 
-    // If FileType couldn't detect the file type (returns null/undefined)
-    if (!fileType) {
-      // For text-based MIME types, accept the declared mimetype since FileType can't detect them
-      if (mimetype && TEXT_MIME_TYPES.includes(mimetype)) {
-        // Optionally verify that the content is valid UTF-8 text
-        try {
-          buffer.toString('utf8');
-          return true;
-        } catch (e) {
-          const err = new Error(`File content is not valid text for mimetype: ${mimetype}`);
-          err.source = "FileContentVerification";
-          throw err;
-        }
-      } else if (mimetype && mimetype.startsWith('image/svg')) {
-        // Handle SVG files (can be image/svg+xml or variants)
-        try {
-          buffer.toString('utf8');
-          return true;
-        } catch (e) {
-          const err = new Error(`File content is not valid text for mimetype: ${mimetype}`);
-          err.source = "FileContentVerification";
-          throw err;
-        }
-      } else {
-        // For non-text files, FileType should be able to detect them
-        const err = new Error(`File content does not match mimetype. Detected: unknown, provided: ${mimetype}`);
+  // If FileType couldn't detect the file type (returns null/undefined)
+  if (!fileType) {
+    // For text-based MIME types, accept the declared mimetype since FileType can't detect them
+    if (mimetype && TEXT_MIME_TYPES.includes(mimetype)) {
+      try {
+        buf.toString('utf8');
+        return true;
+      } catch (e) {
+        const err = new Error(`File content is not valid text for mimetype: ${mimetype}`);
         err.source = "FileContentVerification";
         throw err;
       }
     }
-
-    // If FileType detected a type, it must match the declared mimetype (or be equivalent)
-    if (mimetype && !areMimeTypesEquivalent(fileType.mime, mimetype)) {
-        const err = new Error(`File content does not match mimetype. Detected: ${fileType.mime}, provided: ${mimetype}`);
+    if (mimetype && mimetype.startsWith('image/svg')) {
+      try {
+        buf.toString('utf8');
+        return true;
+      } catch (e) {
+        const err = new Error(`File content is not valid text for mimetype: ${mimetype}`);
         err.source = "FileContentVerification";
         throw err;
+      }
     }
+    if (mimetype && magicMatches(buf, mimetype)) return true;
+    const err = new Error(`File content does not match mimetype. Detected: unknown, provided: ${mimetype}`);
+    err.source = "FileContentVerification";
+    throw err;
+  }
 
-    return true;
-  } catch (err) {
+  // If FileType detected a type, it must match the declared mimetype (or be equivalent)
+  if (mimetype && !areMimeTypesEquivalent(fileType.mime, mimetype)) {
+    const err = new Error(`File content does not match mimetype. Detected: ${fileType.mime}, provided: ${mimetype}`);
+    err.source = "FileContentVerification";
     throw err;
   }
+
+  return true;
 }
 
 module.exports = verifyFileContent;

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@tiledesk/tiledesk-server",
   "description": "The Tiledesk server module",
-  "version": "2.16.0",
+  "version": "2.17.3",
   "scripts": {
     "start": "node ./bin/www",
     "pretest": "mongodb-runner start",

package/routes/filesp.js

@@ -124,9 +124,10 @@ function areMimeTypesEquivalent(mimeType1, mimeType2) {
     'audio/wave': ['audio/wav', 'audio/x-wav', 'audio/vnd.wave'],
     'audio/x-wav': ['audio/wav', 'audio/wave', 'audio/vnd.wave'],
     'audio/vnd.wave': ['audio/wav', 'audio/wave', 'audio/x-wav'],
-    'audio/mpeg': ['audio/opus', 'audio/mp3'],
-    'audio/mp3': ['audio/mpeg', 'audio/opus'],
-    'audio/opus': ['audio/mpeg', 'audio/mp3'],
+    'audio/mpeg': ['audio/opus', 'audio/mp3', 'audio/webm'],
+    'audio/mp3': ['audio/mpeg', 'audio/opus', 'audio/webm'],
+    'audio/opus': ['audio/mpeg', 'audio/mp3', 'audio/webm'],
+    'audio/webm': ['audio/mpeg', 'audio/mp3', 'audio/opus'],
     'image/jpeg': ['image/jpg'],
     'image/jpg': ['image/jpeg'],
     'application/x-zip-compressed': ['application/zip'],

package/routes/kb.js

@@ -3,6 +3,7 @@ var router = express.Router();
 var winston = require('../config/winston');
 var multer = require('multer')
 var upload = multer()
+const path = require('path');
 const JobManager = require('../utils/jobs-worker-queue-manager/JobManagerV2');
 var configGlobal = require('../config/global');
 var mongoose = require('mongoose');
@@ -80,6 +81,37 @@ let default_preview_settings = {
 const default_engine = require('../config/kb/engine');
 const default_engine_hybrid = require('../config/kb/engine.hybrid');
 const default_embedding = require('../config/kb/embedding');
+const PromptManager = require('../config/kb/prompt/rag/PromptManager');
+const situatedContext = require('../config/kb/situatedContext');
+
+const ragPromptManager = new PromptManager(path.join(__dirname, '../config/kb/prompt/rag'));
+
+const RAG_CONTEXT_ENV_OVERRIDES = {
+  "gpt-3.5-turbo":       process.env.GPT_3_5_CONTEXT,
+  "gpt-4":               process.env.GPT_4_CONTEXT,
+  "gpt-4-turbo-preview": process.env.GPT_4T_CONTEXT,
+  "gpt-4o":              process.env.GPT_4O_CONTEXT,
+  "gpt-4o-mini":         process.env.GPT_4O_MINI_CONTEXT,
+  "gpt-4.1":             process.env.GPT_4_1_CONTEXT,
+  "gpt-4.1-mini":        process.env.GPT_4_1_MINI_CONTEXT,
+  "gpt-4.1-nano":        process.env.GPT_4_1_NANO_CONTEXT,
+  "gpt-5":               process.env.GPT_5_CONTEXT,
+  "gpt-5-mini":          process.env.GPT_5_MINI_CONTEXT,
+  "gpt-5-nano":          process.env.GPT_5_NANO_CONTEXT,
+  "general":             process.env.GENERAL_CONTEXT
+};
+
+/** RAG system prompt per modello: file in config/kb/prompt/rag, sovrascrivibili via env (come prima). */
+function getRagContextTemplate(modelName) {
+  const envOverride = RAG_CONTEXT_ENV_OVERRIDES[modelName];
+  if (envOverride) {
+    return envOverride;
+  }
+  if (!PromptManager.modelMap[modelName] && process.env.GENERAL_CONTEXT) {
+    return process.env.GENERAL_CONTEXT;
+  }
+  return ragPromptManager.getPrompt(modelName);
+}
 
 function normalizeEmbedding(embedding) {
   const normalizedEmbedding = (embedding && typeof embedding.toObject === 'function')
@@ -88,19 +120,13 @@ function normalizeEmbedding(embedding) {
   return { ...normalizedEmbedding };
 }
 
-let contexts = {
-  "gpt-3.5-turbo":        process.env.GPT_3_5_CONTEXT       || "You are an helpful assistant for question-answering tasks.\nUse ONLY the pieces of retrieved context delimited by #### and the chat history to answer the question.\nIf you don't know the answer, just say: \"I don't know<NOANS>\"\n\n####{context}####",
-  "gpt-4":                process.env.GPT_4_CONTEXT         || "You are an helpful assistant for question-answering tasks.\nUse ONLY the pieces of retrieved context delimited by #### and the chat history to answer the question.\nIf you don't know the answer, just say that you don't know.\nIf and only if none of the retrieved context is useful for your task, add this word to the end <NOANS>\n\n####{context}####",
-  "gpt-4-turbo-preview":  process.env.GPT_4T_CONTEXT        || "You are an helpful assistant for question-answering tasks.\nUse ONLY the pieces of retrieved context delimited by #### and the chat history to answer the question.\nIf you don't know the answer, just say that you don't know.\nIf and only if none of the retrieved context is useful for your task, add this word to the end <NOANS>\n\n####{context}####",
-  "gpt-4o":               process.env.GPT_4O_CONTEXT        || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, return <NOANS>\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-4o-mini":          process.env.GPT_4O_MINI_CONTEXT   || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, return <NOANS>\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-4.1":              process.env.GPT_4_1_CONTEXT       || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-4.1-mini":         process.env.GPT_4_1_MINI_CONTEXT  || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-4.1-nano":         process.env.GPT_4_1_NANO_CONTEXT  || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-5":                process.env.GPT_5_CONTEXT         || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-5-mini":           process.env.GPT_5_MINI_CONTEXT    || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "gpt-5-nano":           process.env.GPT_5_NANO_CONTEXT    || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end==",
-  "general":              process.env.GENERAL_CONTEXT       || "You are an helpful assistant for question-answering tasks. Follow these steps carefully:\n1. Answer in the same language of the user question, regardless of the retrieved context language\n2. Use ONLY the pieces of the retrieved context and the chat history to answer the question.\n3. If the retrieved context does not contain sufficient information to generate an accurate and informative answer, append <NOANS> at the end of the answer\n\n==Retrieved context start==\n{context}\n==Retrieved context end=="
+function normalizeSituatedContext() {
+  return situatedContext.enable
+    ? {
+      ...situatedContext,
+      api_key: process.env.SITUATED_CONTEXT_API_KEY || process.env.GPTKEY
+    }
+    : undefined;
 }
 
 /**
@@ -236,6 +262,11 @@ router.post('/scrape/single', async (req, res) => {
         json.hybrid = true;
       }
 
+      const situated_context = normalizeSituatedContext();
+      if (situated_context) {
+        json.situated_context = situated_context;
+      }
+
       winston.verbose("/scrape/single json: ", json);
 
       if (process.env.NODE_ENV === "test") {
@@ -361,7 +392,7 @@ router.post('/qa', async (req, res) => {
 
   // Check if "Advanced Mode" is active. In such case the default_context must be not appended
   if (!data.advancedPrompt) {
-    const contextTemplate = contexts[data.model.name] || contexts["general"];
+    const contextTemplate = getRagContextTemplate(data.model.name);
     if (data.system_context) {
       data.system_context = data.system_context + " \n" + contextTemplate;
     } else {
@@ -393,7 +424,7 @@ router.post('/qa', async (req, res) => {
     }
   }
 
-  data.stream = false;
+  data.stream = data.stream === true;
   data.debug = true;
   delete data.advancedPrompt;
   winston.verbose("ask data: ", data);
@@ -402,16 +433,163 @@ router.post('/qa', async (req, res) => {
     return res.status(200).send({ success: true, message: "Question skipped in test environment", data: data });
   }
 
+  if (data.stream === true) {
+    // Streaming SSE: use askStream and forward only content as JSON SSE events
+    res.status(200);
+    res.setHeader('Content-Type', 'text/event-stream');
+    res.setHeader('Cache-Control', 'no-cache');
+    res.setHeader('Connection', 'keep-alive');
+    res.setHeader('Access-Control-Allow-Origin', '*');
+
+    const sendError = (message) => {
+      try {
+        res.write('data: ' + JSON.stringify({ error: message }) + '\n\n');
+      } catch (_) {}
+      res.end();
+    };
+
+    function extractContent(obj) {
+      if (obj.content != null) return obj.content;
+      if (obj.choices && obj.choices[0]) {
+        const c = obj.choices[0];
+        if (c.delta && c.delta.content != null) return c.delta.content;
+        if (c.message && c.message.content != null) return c.message.content;
+      }
+      return null;
+    }
+
+    /** Same JSON shape as non-stream /qa: stream may wrap it in model_used */
+    function normalizeKbQaPayload(obj) {
+      if (obj && typeof obj === 'object' && obj.model_used != null && typeof obj.model_used === 'object') {
+        return obj.model_used;
+      }
+      return obj;
+    }
+
+    /** Flat final payload like non-stream /qa (answer, prompt_token_size, …) */
+    function isMetadataPayload(obj, streamedContent) {
+      if (obj == null || typeof obj !== 'object') return false;
+      if (streamedContent != null && streamedContent !== '') return false;
+      if (typeof obj.prompt_token_size === 'number') return true;
+      if (obj.answer != null) return true;
+      if (obj.sources != null) return true;
+      if (obj.chunks != null) return true;
+      if (obj.content_chunks != null) return true;
+      return false;
+    }
+
+    /** KB stream summary: full_response + model_used (same info as non-stream body, plus envelope) */
+    function isKbStreamCompletedSummary(obj) {
+      if (obj == null || typeof obj !== 'object') return false;
+      if (obj.status === 'completed') return true;
+      if (obj.full_response != null && obj.model_used != null && typeof obj.model_used === 'object') return true;
+      return false;
+    }
+
+    function forwardSsePayload(payload) {
+      if (payload === '[DONE]') return;
+      let obj;
+      try {
+        obj = JSON.parse(payload);
+      } catch (_) {
+        return;
+      }
+
+      if (obj.status === 'started') {
+        return;
+      }
+      if (isKbStreamCompletedSummary(obj)) {
+        res.write('data: ' + JSON.stringify(normalizeKbQaPayload(obj)) + '\n\n');
+        return;
+      }
+
+      if (obj.type === 'metadata' || obj.event === 'metadata') {
+        res.write('data: ' + JSON.stringify(normalizeKbQaPayload(obj)) + '\n\n');
+        return;
+      }
+      const content = extractContent(obj);
+      if (content != null && content !== '') {
+        res.write('data: ' + JSON.stringify({ content }) + '\n\n');
+        return;
+      }
+      const normalized = normalizeKbQaPayload(obj);
+      if (isMetadataPayload(normalized, content)) {
+        res.write('data: ' + JSON.stringify(normalized) + '\n\n');
+      }
+    }
+
+    aiService.askStream(data).then((resp) => {
+      const stream = resp.data;
+      let buffer = '';
+
+      stream.on('data', (chunk) => {
+        buffer += chunk.toString();
+        const lines = buffer.split('\n');
+        buffer = lines.pop() || '';
+
+        for (const line of lines) {
+          const trimmed = line.trim();
+          if (!trimmed.startsWith('data: ')) continue;
+          const payload = trimmed.slice(6);
+          forwardSsePayload(payload);
+        }
+      });
+
+      stream.on('end', () => {
+        const tail = buffer.trim();
+        if (tail) {
+          for (const line of tail.split('\n')) {
+            const trimmed = line.trim();
+            if (!trimmed.startsWith('data: ')) continue;
+            forwardSsePayload(trimmed.slice(6));
+          }
+        }
+        res.write('data: [DONE]\n\n');
+        res.end();
+      });
+
+      stream.on('error', (err) => {
+        winston.error('qa stream err: ', err);
+        sendError(err.message || 'Stream error');
+      });
+
+      res.on('close', () => {
+        if (!res.writableEnded) {
+          stream.destroy();
+        }
+      });
+    }).catch((err) => {
+      winston.error('qa err: ', err);
+      winston.error('qa err.response: ', err.response);
+      const message = (err.response && err.response.data && typeof err.response.data.pipe !== 'function' && err.response.data.detail)
+        ? err.response.data.detail
+        : (err.response && err.response.statusText) || err.message || String(err);
+      if (!res.headersSent) {
+        res.status(err.response && err.response.status ? err.response.status : 500);
+      }
+      sendError(message);
+    });
+    return;
+  }
+
   aiService.askNamespace(data).then((resp) => {
     winston.debug("qa resp: ", resp.data);
     let answer = resp.data;
 
     if (publicKey === true) {
-      let multiplier = MODELS_MULTIPLIER[data.model];
+      let modelKey;
+      if (typeof data.model === 'string') {
+        modelKey = data.model;
+      } else if (data.model && typeof data.model.name === 'string') {
+        modelKey = data.model.name;
+      }
+
+      let multiplier = MODELS_MULTIPLIER[modelKey];
       if (!multiplier) {
         multiplier = 1;
-        winston.info("No multiplier found for AI model (qa) " + data.model);
+        winston.info("No multiplier found for AI model (qa) " + modelKey);
       }
+
       obj.multiplier = multiplier;
       obj.tokens = answer.prompt_token_size;
 
@@ -1039,6 +1217,7 @@ router.post('/namespace/import/:id', upload.single('uploadFile'), async (req, re
   let embedding = normalizeEmbedding(ns.embedding);
   embedding.api_key = process.env.EMBEDDING_API_KEY || process.env.GPTKEY;
   let hybrid = ns.hybrid;
+  const situated_context = normalizeSituatedContext();
 
 
   if (process.env.NODE_ENV !== "test") {
@@ -1076,7 +1255,13 @@ router.post('/namespace/import/:id', upload.single('uploadFile'), async (req, re
 
   let resources = new_contents.map(({ name, status, __v, createdAt, updatedAt, id_project, ...keepAttrs }) => keepAttrs)
   resources = resources.map(({ _id, scrape_options, ...rest }) => {
-    return { id: _id, parameters_scrape_type_4: scrape_options, embedding: embedding, engine: engine, ...rest}
+    return { 
+      id: _id, 
+      parameters_scrape_type_4: scrape_options, 
+      embedding: embedding, 
+      engine: engine, 
+      ...(situated_context && { situated_context: situated_context }), 
+      ...rest}
   });
   
   winston.verbose("resources to be sent to worker: ", resources);
@@ -1420,13 +1605,14 @@ router.post('/', async (req, res) => {
   }
   if (type === 'url') {
     new_kb.refresh_rate = refresh_rate || 'never';
-    if (!scrape_type || scrape_type === 2) {
-      new_kb.scrape_type = 2;
-      new_kb.scrape_options = aiManager.setDefaultScrapeOptions();
-    } else {
+    if (scrape_type === 0 || scrape_type === 4) {
       new_kb.scrape_type = scrape_type;
       new_kb.scrape_options = scrape_options;
     }
+    else {
+      new_kb.scrape_type = 2;
+      new_kb.scrape_options = aiManager.setDefaultScrapeOptions();
+    }
   }
 
   if (tags && Array.isArray(tags) && tags.every(tag => typeof tag === "string")) {
@@ -1451,6 +1637,8 @@ router.post('/', async (req, res) => {
       const embedding = normalizeEmbedding(namespace.embedding);
       embedding.api_key = process.env.EMBEDDING_API_KEY || process.env.GPTKEY;
 
+      const situated_context = normalizeSituatedContext();
+
       const json = {
         id: saved_kb._id,
         type: saved_kb.type,
@@ -1461,6 +1649,7 @@ router.post('/', async (req, res) => {
         hybrid: namespace.hybrid,
         engine: namespace.engine || default_engine,
         embedding: embedding,
+        ...(situated_context && { situated_context: situated_context }),
         ...(saved_kb.scrape_type && { scrape_type: saved_kb.scrape_type }),
         ...(saved_kb.scrape_options && { parameters_scrape_type_4: saved_kb.scrape_options }),
         ...(saved_kb.tags && { tags: saved_kb.tags }),
@@ -1617,10 +1806,18 @@ router.post('/csv', upload.single('uploadFile'), async (req, res) => {
         let embedding = normalizeEmbedding(namespace.embedding);
         embedding.api_key = process.env.EMBEDDING_API_KEY || process.env.GPTKEY;
         let hybrid = namespace.hybrid;
+        const situated_context = normalizeSituatedContext();
 
         let resources = result.map(({ name, status, __v, createdAt, updatedAt, id_project,  ...keepAttrs }) => keepAttrs)
         resources = resources.map(({ _id, ...rest}) => {
-          return { id: _id, webhook: webhook, embedding: embedding, engine: engine, ...rest };
+          return { 
+            id: _id, 
+            webhook: webhook, 
+            embedding: embedding, 
+            engine: engine, 
+            ...(situated_context && { situated_context: situated_context }), 
+            ...rest 
+          };
         })
         winston.verbose("resources to be sent to worker: ", resources);
 
@@ -1856,13 +2053,14 @@ router.put('/:kb_id', async (req, res) => {
 
   if (new_content.type === 'url') {
     new_content.refresh_rate = refresh_rate || 'never';
-    if (!scrape_type || scrape_type === 2) {
-      new_content.scrape_type = 2;
-      new_content.scrape_options = aiManager.setDefaultScrapeOptions();
-    } else {
+    if (scrape_type === 0 || scrape_type === 4) {
       new_content.scrape_type = scrape_type;
       new_content.scrape_options = scrape_options;
     }
+    else {
+      new_content.scrape_type = 2;
+      new_content.scrape_options = aiManager.setDefaultScrapeOptions();
+    }
   }
 
   if (kb.sitemap_origin_id) {
@@ -1887,6 +2085,7 @@ router.put('/:kb_id', async (req, res) => {
   const embedding = normalizeEmbedding(namespace.embedding);
   embedding.api_key = process.env.EMBEDDING_API_KEY || process.env.GPTKEY;
   let webhook = apiUrl + '/webhook/kb/status?token=' + KB_WEBHOOK_TOKEN;
+  const situated_context = normalizeSituatedContext();
 
   const json = {
     id: updated_content._id,
@@ -1898,6 +2097,7 @@ router.put('/:kb_id', async (req, res) => {
     hybrid: namespace.hybrid,
     engine: namespace.engine || default_engine,
     embedding: embedding,
+    ...(situated_context && { situated_context: situated_context }),
     ...(updated_content.scrape_type && { scrape_type: updated_content.scrape_type }),
     ...(updated_content.scrape_options && { parameters_scrape_type_4: updated_content.scrape_options }),
     ...(updated_content.tags && { tags: updated_content.tags }),
@@ -1914,40 +2114,6 @@ router.put('/:kb_id', async (req, res) => {
 
 })
 
-// router.put('/:kb_id', async (req, res) => {
-
-//   let kb_id = req.params.kb_id;
-//   winston.verbose("update kb_id " + kb_id);
-
-//   let update = {};
-
-//   if (req.body.name != undefined) {
-//     update.name = req.body.name;
-//   }
-
-//   if (req.body.status != undefined) {
-//     update.status = req.body.status;
-//   }
-
-//   winston.debug("kb update: ", update);
-
-//   KB.findByIdAndUpdate(kb_id, update, { new: true }, (err, savedKb) => {
-
-//     if (err) {
-//       winston.error("KB findByIdAndUpdate error: ", err);
-//       return res.status(500).send({ success: false, error: err });
-//     }
-
-//     if (!savedKb) {
-//       winston.debug("Try to updating a non-existing kb");
-//       return res.status(400).send({ success: false, message: "Content not found" })
-//     }
-
-//     res.status(200).send(savedKb)
-//   })
-
-// })
-
 router.delete('/:kb_id', async (req, res) => {
 
   let project_id = req.projectid;

package/routes/quotes.js

@@ -44,10 +44,17 @@ router.post('/incr/:type', async (req, res) => {
 
     let quoteManager = req.app.get('quote_manager');
 
-    let multiplier = MODELS_MULTIPLIER[data.model];
+    let modelKey;
+    if (typeof data.model === 'string') {
+        modelKey = data.model;
+    } else if (data.model && typeof data.model.name === 'string') {
+        modelKey = data.model.name;
+    }
+
+    let multiplier = MODELS_MULTIPLIER[modelKey];
     if (!multiplier) {
         multiplier = 1;
-        winston.info("No multiplier found for AI model (incr) " + data.model)
+        winston.info("No multiplier found for AI model (incr) " + modelKey)
     }
     data.multiplier = multiplier;
     data.createdAt = new Date();

package/routes/webhook.js

@@ -193,6 +193,11 @@ router.post('/kb/reindex', async (req, res) => {
     embedding.api_key = process.env.EMBEDDING_API_KEY || process.env.GPTKEY;
     json.embedding = embedding;
 
+    const situated_context = aiManager.normalizeSituatedContext();
+    if (situated_context) {
+      json.situated_context = situated_context;
+    }
+
     let resources = [];
     resources.push(json);
   

package/services/aiManager.js

@@ -20,6 +20,7 @@ const default_engine = require('../config/kb/engine');
 const default_engine_hybrid = require('../config/kb/engine.hybrid');
 const default_embedding = require('../config/kb/embedding');
 const integrationService = require('./integrationService');
+const situatedContext = require('../config/kb/situatedContext');
 
 // Job managers
 let jobManager = new JobManager(AMQP_MANAGER_URL, {
@@ -93,11 +94,22 @@ class AiManager {
         let engine = namespace.engine || default_engine;
         let embedding = namespace.embedding || default_embedding;
         embedding.api_key = process.env.EMBEDDING_API_KEY || process.env.GPTKEY;
+
+        let situated_context = this.normalizeSituatedContext();
+
         let webhook = apiUrl + '/webhook/kb/status?token=' + KB_WEBHOOK_TOKEN;
 
         let resources = result.map(({ name, status, __v, createdAt, updatedAt, id_project, ...keepAttrs }) => keepAttrs)
         resources = resources.map(({ _id, scrape_options, ...rest }) => {
-          return { id: _id, webhook: webhook, parameters_scrape_type_4: scrape_options, embedding: embedding, engine: engine, hybrid: hybrid, ...rest}
+          return { 
+            id: _id, 
+            webhook: webhook, 
+            parameters_scrape_type_4: scrape_options, 
+            embedding: embedding, 
+            engine: engine, 
+            hybrid: hybrid, 
+            ...(situated_context && { situated_context }),
+            ...rest}
         });
 
         winston.verbose("resources to be sent to worker: ", resources);
@@ -120,6 +132,8 @@ class AiManager {
   async scheduleSitemap(namespace, sitemap_content, options) {
     return new Promise((resolve, reject) => {
 
+      const situated_context = this.normalizeSituatedContext();
+
       let kb = {
         id: sitemap_content._id,
         source: sitemap_content.source,
@@ -130,6 +144,7 @@ class AiManager {
         engine: namespace.engine,
         embedding: namespace.embedding,
         hybrid: namespace.hybrid,
+        ...(situated_context && { situated_context }),
       }
 
       if (process.env.NODE_ENV === 'test') {
@@ -552,6 +567,15 @@ class AiManager {
     })
   }
 
+  normalizeSituatedContext() {
+    return situatedContext.enable
+      ? {
+        ...situatedContext,
+        api_key: process.env.SITUATED_CONTEXT_API_KEY || process.env.GPTKEY
+      }
+      : undefined;
+  }
+
 }
 
 const aiManager = new AiManager();

package/services/aiService.js

@@ -206,16 +206,18 @@ class AiService {
     }
     winston.debug("[OPENAI SERVICE] kb endpoint: " + base_url);
 
+    const config = {
+      url: base_url + "/qa",
+      headers: {
+        'Content-Type': 'application/json'
+      },
+      data: data,
+      method: 'POST'
+    };
+
     return new Promise((resolve, reject) => {
 
-      axios({
-        url: base_url + "/qa",
-        headers: {
-          'Content-Type': 'application/json'
-        },
-        data: data,
-        method: 'POST'
-      }).then((resbody) => {
+      axios(config).then((resbody) => {
         resolve(resbody);
       }).catch((err) => {
         reject(err);
@@ -224,6 +226,29 @@ class AiService {
     })
   }
 
+  /**
+   * Stream /qa from KB service. Uses Axios with responseType: 'stream'.
+   * Returns the raw Axios response (resp.data is the Node.js Readable stream).
+   */
+  askStream(data) {
+    winston.debug("askStream data: ", data);
+    let base_url = kb_endpoint_qa;
+    if (data.hybrid || data.search_type === 'hybrid') {
+      base_url = kb_endpoint_qa_gpu;
+    }
+    winston.debug("[OPENAI SERVICE] kb stream endpoint: " + base_url);
+
+    return axios({
+      url: base_url + "/qa",
+      headers: {
+        'Content-Type': 'application/json'
+      },
+      data: data,
+      method: 'POST',
+      responseType: 'stream'
+    });
+  }
+
   getContentChunks(namespace_id, content_id, engine, hybrid) {
     let base_url = kb_endpoint_train;
     winston.debug("[OPENAI SERVICE] kb endpoint: " + base_url);

package/services/fileGridFsService.js

@@ -177,10 +177,9 @@ class FileGridFsService extends FileService {
                 return reject(e);
             })
             stream.on('data', (data) => {
-                bufs.push(data);
+                bufs.push(Buffer.isBuffer(data) ? data : Buffer.from(data));
             });
             stream.on('end', () => {
-
                 var buffer = Buffer.concat(bufs);
                 return resolve(buffer);
             });