@exulu/backend 1.50.0 → 1.51.0

package/dist/index.cjs

@@ -205,17 +205,19 @@ async function postgresClient() {
           // 30 minutes
         },
         pool: {
-          min: 5,
-          // Increased from 2 to ensure enough connections available
-          max: 50,
-          // Increased from 20 to handle more concurrent operations with processor jobs
-          acquireTimeoutMillis: 6e4,
-          // Increased from 30s to 60s to handle pool contention
+          min: 10,
+          // Minimum connections always ready
+          max: 300,
+          // Increased to support high worker concurrency (250+ concurrent jobs)
+          acquireTimeoutMillis: 12e4,
+          // 2 minutes - increased to handle high contention during bursts
           createTimeoutMillis: 3e4,
           idleTimeoutMillis: 6e4,
-          // Increased to keep connections alive longer
+          // Keep connections alive for reuse
           reapIntervalMillis: 1e3,
           createRetryIntervalMillis: 200,
+          // Enable propagateCreateError to properly handle connection creation failures
+          propagateCreateError: false,
           // Log pool events to help debug connection issues
           afterCreate: (conn, done) => {
             console.log("[EXULU] New database connection created");
@@ -3618,7 +3620,7 @@ var convertExuluToolsToAiSdkTools = async (currentTools, approvedTools, allExulu
           description,
           // The approvedTools array uses the tool.name lookup as the frontend
           // Vercel AI SDK uses the sanitized tool name as the key, so this matches.
-          needsApproval: approvedTools?.includes("tool-" + cur.name) ? false : true,
+          needsApproval: approvedTools?.includes("tool-" + cur.name) || !cur.needsApproval ? false : true,
           // todo make configurable
           async *execute(inputs, options) {
             console.log(
@@ -3768,6 +3770,7 @@ var ExuluTool = class {
   inputSchema;
   type;
   tool;
+  needsApproval;
   config;
   constructor({
     id,
@@ -3777,10 +3780,12 @@ var ExuluTool = class {
     inputSchema,
     type,
     execute: execute2,
-    config
+    config,
+    needsApproval
   }) {
     this.id = id;
     this.config = config;
+    this.needsApproval = needsApproval ?? true;
     this.category = category || "default";
     this.name = name;
     this.description = description;
@@ -4187,11 +4192,30 @@ var ExuluContext2 = class {
     );
     await db2.from(getChunksTableName(this.id)).where({ source }).delete();
     if (chunks?.length) {
+      const sanitizeString = (str) => {
+        if (!str) return "";
+        return str.replace(/\0/g, "");
+      };
+      const sanitizeMetadata2 = (metadata) => {
+        if (!metadata) return {};
+        const sanitized = {};
+        for (const [key, value] of Object.entries(metadata)) {
+          if (typeof value === "string") {
+            sanitized[key] = sanitizeString(value);
+          } else {
+            sanitized[key] = value;
+          }
+        }
+        return sanitized;
+      };
       await db2.from(getChunksTableName(this.id)).insert(
         chunks.map((chunk) => ({
-          source,
-          metadata: chunk.metadata,
-          content: chunk.content,
+          // Sanitize source to remove null bytes
+          source: sanitizeString(source),
+          // Sanitize metadata to remove null bytes from string values
+          metadata: sanitizeMetadata2(chunk.metadata),
+          // Remove null bytes (0x00) which are invalid in PostgreSQL UTF8 encoding
+          content: sanitizeString(chunk.content),
           chunk_index: chunk.index,
           embedding: import_knex5.default.toSql(chunk.vector)
         }))
@@ -4580,6 +4604,8 @@ var ExuluContext2 = class {
       name: `${this.name}_context_search`,
       type: "context",
       category: "contexts",
+      needsApproval: true,
+      // todo make configurable
       inputSchema: import_zod4.z.object({
         query: import_zod4.z.string().describe("The original question that the user asked"),
         keywords: import_zod4.z.array(import_zod4.z.string()).describe(
@@ -5773,7 +5799,7 @@ var finalizeRequestedFields = async ({
           return result;
         }
         const { db: db2 } = await postgresClient();
-        const query = db2.from(getChunksTableName(context.id)).where({ source: result.id }).select("id", "content", "source", "chunk_index", "createdAt", "updatedAt");
+        const query = db2.from(getChunksTableName(context.id)).where({ source: result.id }).select("id", "content", "source", "chunk_index", "createdAt", "updatedAt", "metadata");
         const chunks = await query;
         result.chunks = chunks.map((chunk) => ({
           chunk_content: chunk.content,
@@ -5786,7 +5812,8 @@ var finalizeRequestedFields = async ({
           item_created_at: chunk.item_created_at,
           item_id: chunk.item_id,
           item_external_id: chunk.item_external_id,
-          item_name: chunk.item_name
+          item_name: chunk.item_name,
+          chunk_metadata: chunk.metadata
         }));
       }
     }
@@ -7160,6 +7187,36 @@ var import_ai3 = require("ai");
 var import_crypto_js4 = __toESM(require("crypto-js"), 1);
 var redisConnection;
 var unhandledRejectionHandlerInstalled = false;
+var poolMonitoringInterval;
+var startPoolMonitoring = () => {
+  if (poolMonitoringInterval) return;
+  poolMonitoringInterval = setInterval(async () => {
+    try {
+      const { db: db2 } = await postgresClient();
+      const poolStats = db2.client.pool;
+      if (poolStats) {
+        const used = poolStats.numUsed?.() || 0;
+        const free = poolStats.numFree?.() || 0;
+        const pending = poolStats.numPendingAcquires?.() || 0;
+        const total = used + free;
+        console.log("[EXULU] Connection pool health check:", {
+          used,
+          free,
+          pending,
+          total,
+          utilization: total > 0 ? `${Math.round(used / total * 100)}%` : "0%"
+        });
+        if (pending > 10) {
+          console.warn(
+            `[EXULU] WARNING: ${pending} jobs waiting for database connections. Consider increasing pool size or reducing worker concurrency.`
+          );
+        }
+      }
+    } catch (error) {
+      console.error("[EXULU] Error checking pool health:", error);
+    }
+  }, 3e4);
+};
 var installGlobalErrorHandlers = () => {
   if (unhandledRejectionHandlerInstalled) return;
   process.on("unhandledRejection", (reason) => {
@@ -7184,6 +7241,7 @@ var installGlobalErrorHandlers = () => {
   unhandledRejectionHandlerInstalled = true;
   console.log("[EXULU] Global error handlers installed to prevent worker crashes");
 };
+var isShuttingDown = false;
 var createWorkers = async (providers, queues2, config, contexts, rerankers, evals, tools, tracer) => {
   console.log("[EXULU] creating workers for " + queues2?.length + " queues.");
   console.log(
@@ -7191,7 +7249,8 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
     queues2.map((q) => q.queue.name)
   );
   installGlobalErrorHandlers();
-  process.setMaxListeners(Math.max(queues2.length * 2 + 5, 15));
+  startPoolMonitoring();
+  process.setMaxListeners(Math.max(15, process.getMaxListeners()));
   if (!redisServer.host || !redisServer.port) {
     console.error(
       "[EXULU] you are trying to start worker, but no redis server is configured in the environment."
@@ -7224,7 +7283,53 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
           status: await bullmqJob.getState(),
           type: bullmqJob.data.type
         });
-        const { db: db2 } = await postgresClient();
+        let progressInterval;
+        if (bullmqJob.data.type === "processor") {
+          progressInterval = setInterval(async () => {
+            try {
+              await bullmqJob.updateProgress({
+                status: "processing",
+                timestamp: (/* @__PURE__ */ new Date()).toISOString()
+              });
+              console.log(`[EXULU] Job ${bullmqJob.id} heartbeat sent to prevent stalling`);
+            } catch (error) {
+              console.error(`[EXULU] Error updating job progress:`, error);
+            }
+          }, 25e3);
+        }
+        let db2;
+        let retries = 3;
+        let lastError;
+        for (let attempt = 1; attempt <= retries; attempt++) {
+          try {
+            const client2 = await postgresClient();
+            db2 = client2.db;
+            const poolStats = db2.client.pool;
+            if (poolStats) {
+              console.log(`[EXULU] Connection pool stats for job ${bullmqJob.id}:`, {
+                size: poolStats.numUsed?.() || 0,
+                available: poolStats.numFree?.() || 0,
+                pending: poolStats.numPendingAcquires?.() || 0
+              });
+            }
+            break;
+          } catch (error) {
+            lastError = error instanceof Error ? error : new Error(String(error));
+            console.error(
+              `[EXULU] Failed to acquire database connection (attempt ${attempt}/${retries}) for job ${bullmqJob.id}:`,
+              lastError.message
+            );
+            if (attempt < retries) {
+              const backoffMs = 500 * Math.pow(2, attempt - 1);
+              await new Promise((resolve3) => setTimeout(resolve3, backoffMs));
+            }
+          }
+        }
+        if (!db2) {
+          throw new Error(
+            `Failed to acquire database connection after ${retries} attempts: ${lastError?.message}`
+          );
+        }
         const data = bullmqJob.data;
         const timeoutInSeconds = data.timeoutInSeconds || queue.timeoutInSeconds || 600;
         const timeoutMs = timeoutInSeconds * 1e3;
@@ -7316,7 +7421,7 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
               }
               const exuluStorage = new ExuluStorage({ config });
               console.log("[EXULU] POS 2 -- EXULU CONTEXT PROCESS FIELD");
-              const processorResult = await context.processor.execute({
+              let processorResult = await context.processor.execute({
                 item: data.inputs,
                 user: data.user,
                 role: data.role,
@@ -7331,12 +7436,16 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
                 );
               }
               delete processorResult.field;
+              const updateData = { ...processorResult };
               await db2.from(getTableName(context.id)).where({
                 id: processorResult.id
               }).update({
-                ...processorResult,
+                ...updateData,
                 last_processed_at: (/* @__PURE__ */ new Date()).toISOString()
               });
+              Object.keys(updateData).forEach((key) => {
+                delete updateData[key];
+              });
               let jobs = [];
               if (context.processor?.config?.generateEmbeddings) {
                 const fullItem = await db2.from(getTableName(context.id)).where({
@@ -7358,12 +7467,18 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
                   jobs.push(embeddingsJob);
                 }
               }
-              return {
-                result: processorResult,
+              const result = {
+                result: { id: processorResult.id },
                 metadata: {
                   jobs: jobs.length > 0 ? jobs.join(",") : void 0
                 }
               };
+              processorResult = null;
+              const memUsage = process.memoryUsage();
+              console.log(
+                `[EXULU] Memory after processor job ${bullmqJob.id}: ${Math.round(memUsage.heapUsed / 1024 / 1024)}MB / ${Math.round(memUsage.heapTotal / 1024 / 1024)}MB`
+              );
+              return result;
             }
             if (data.type === "workflow") {
               console.log("[EXULU] running a workflow job.", bullmqJob.name);
@@ -7382,10 +7497,10 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
                 user,
                 messages: inputMessages
               } = await validateWorkflowPayload(data, providers);
-              const retries = 3;
+              const retries2 = 3;
               let attempts = 0;
               const promise = new Promise(async (resolve3, reject) => {
-                while (attempts < retries) {
+                while (attempts < retries2) {
                   try {
                     const messages2 = await processUiMessagesFlow({
                       providers,
@@ -7407,7 +7522,7 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
                       error instanceof Error ? error.message : String(error)
                     );
                     attempts++;
-                    if (attempts >= retries) {
+                    if (attempts >= retries2) {
                       reject(new Error(error instanceof Error ? error.message : String(error)));
                     }
                     await new Promise((resolve4) => setTimeout((resolve5) => resolve5(true), 2e3));
@@ -7458,10 +7573,10 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
                 testCase,
                 messages: inputMessages
               } = await validateEvalPayload(data, providers);
-              const retries = 3;
+              const retries2 = 3;
               let attempts = 0;
               const promise = new Promise(async (resolve3, reject) => {
-                while (attempts < retries) {
+                while (attempts < retries2) {
                   try {
                     const messages2 = await processUiMessagesFlow({
                       providers,
@@ -7482,7 +7597,7 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
                       error instanceof Error ? error.message : String(error)
                     );
                     attempts++;
-                    if (attempts >= retries) {
+                    if (attempts >= retries2) {
                       reject(new Error(error instanceof Error ? error.message : String(error)));
                     }
                     await new Promise((resolve4) => setTimeout((resolve5) => resolve5(true), 2e3));
@@ -7731,9 +7846,15 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
         try {
           const result = await Promise.race([workPromise, timeoutPromise]);
           clearTimeout(timeoutHandle);
+          if (progressInterval) {
+            clearInterval(progressInterval);
+          }
           return result;
         } catch (error) {
           clearTimeout(timeoutHandle);
+          if (progressInterval) {
+            clearInterval(progressInterval);
+          }
           console.error(
             `[EXULU] job ${bullmqJob.id} failed (error caught in race handler).`,
             error instanceof Error ? error.message : String(error)
@@ -7747,6 +7868,16 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
         concurrency: queue.concurrency?.worker || 1,
         removeOnComplete: { count: 1e3 },
         removeOnFail: { count: 5e3 },
+        // Configure settings for long-running jobs (especially processor jobs)
+        // lockDuration: How long a worker can hold a job before it's considered stalled
+        // Set to 5 minutes to accommodate CPU-intensive operations
+        lockDuration: 3e5,
+        // 5 minutes in milliseconds
+        // stalledInterval: How often to check for stalled jobs
+        // Set to 2 minutes to reduce false positives for long-running operations
+        stalledInterval: 12e4,
+        // 2 minutes in milliseconds
+        maxStalledCount: 1,
         ...queue.ratelimit && {
           limiter: {
             max: queue.ratelimit,
@@ -7783,24 +7914,68 @@ var createWorkers = async (providers, queues2, config, contexts, rerankers, eval
           error: error instanceof Error ? error.message : String(error)
         } : error
       );
+      throw error;
     });
     worker.on("error", (error) => {
       console.error(`[EXULU] worker error.`, error);
+      throw error;
     });
     worker.on("progress", (job, progress) => {
       console.log(`[EXULU] job progress ${job.id}.`, job.name, {
         progress
       });
     });
-    const gracefulShutdown = async (signal) => {
-      console.log(`Received ${signal}, closing server...`);
-      await worker.close();
-      process.exit(0);
-    };
-    process.on("SIGINT", () => gracefulShutdown("SIGINT"));
-    process.on("SIGTERM", () => gracefulShutdown("SIGTERM"));
     return worker;
   });
+  const gracefulShutdown = async (signal) => {
+    if (isShuttingDown) {
+      console.log(`[EXULU] Shutdown already in progress, ignoring additional ${signal}`);
+      return;
+    }
+    isShuttingDown = true;
+    console.log(`[EXULU] Received ${signal}, shutting down gracefully...`);
+    try {
+      if (poolMonitoringInterval) {
+        clearInterval(poolMonitoringInterval);
+        poolMonitoringInterval = void 0;
+      }
+      console.log(`[EXULU] Closing ${workers.length} worker(s)...`);
+      const closePromises = workers.map(async (worker, index) => {
+        try {
+          await Promise.race([
+            worker.close(),
+            new Promise(
+              (_, reject) => setTimeout(() => reject(new Error("Worker close timeout")), 3e4)
+            )
+          ]);
+          console.log(`[EXULU] Worker ${index + 1} closed successfully`);
+        } catch (error) {
+          console.error(`[EXULU] Error closing worker ${index + 1}:`, error);
+        }
+      });
+      await Promise.allSettled(closePromises);
+      if (redisConnection) {
+        console.log(`[EXULU] Closing Redis connection...`);
+        await redisConnection.quit();
+      }
+      try {
+        const { db: db2 } = await postgresClient();
+        if (db2?.client) {
+          console.log(`[EXULU] Closing database connection pool...`);
+          await db2.client.destroy();
+        }
+      } catch (error) {
+        console.error(`[EXULU] Error closing database:`, error);
+      }
+      console.log(`[EXULU] Graceful shutdown complete`);
+      process.exit(0);
+    } catch (error) {
+      console.error(`[EXULU] Error during graceful shutdown:`, error);
+      process.exit(1);
+    }
+  };
+  process.once("SIGINT", () => gracefulShutdown("SIGINT"));
+  process.once("SIGTERM", () => gracefulShutdown("SIGTERM"));
   return workers;
 };
 var validateWorkflowPayload = async (data, providers) => {
@@ -9514,6 +9689,7 @@ type AgentEvalFunctionConfig {
 
 type ItemChunks {
     chunk_id: String!
+    chunk_metadata: JSON!
     chunk_index: Int!
     chunk_content: String!
     chunk_source: String!
@@ -9920,7 +10096,7 @@ var ExuluProvider = class {
         prompt: import_zod7.z.string().describe("The prompt (usually a question for the agent) to send to the agent."),
         information: import_zod7.z.string().describe("A summary of relevant context / information from the current session")
       }),
-      description: `This tool calls an AI agent named: ${agent.name}. The agent does the following: ${agent.description}.`,
+      description: `This tool calls an agent named: ${agent.name}. The agent does the following: ${agent.description}.`,
       config: [],
       execute: async ({ prompt, information, user, allExuluTools }) => {
         const hasAccessToAgent = await checkRecordAccess(agent, "read", user);
@@ -10033,9 +10209,6 @@ var ExuluProvider = class {
     if (!prompt && !inputMessages?.length) {
       throw new Error("Prompt or message is required for generating.");
     }
-    if (outputSchema && !prompt) {
-      throw new Error("Prompt is required for generating with an output schema.");
-    }
     const model = this.model.create({
       ...providerapikey ? { apiKey: providerapikey } : {}
     });
@@ -10172,14 +10345,18 @@ var ExuluProvider = class {
       let inputTokens = 0;
       let outputTokens = 0;
       if (outputSchema) {
-        const { object, usage } = await (0, import_ai4.generateObject)({
+        const { output, usage } = await (0, import_ai4.generateText)({
           model,
           system,
-          prompt,
           maxRetries: 3,
-          schema: outputSchema
+          output: import_ai4.Output.object({
+            schema: outputSchema
+          }),
+          prompt,
+          stopWhen: [(0, import_ai4.stepCountIs)(5)]
+          // make configurable
         });
-        result.object = object;
+        result.object = output;
         inputTokens = usage.inputTokens || 0;
         outputTokens = usage.outputTokens || 0;
       } else {
@@ -10210,6 +10387,7 @@ var ExuluProvider = class {
             agent
           ),
           stopWhen: [(0, import_ai4.stepCountIs)(5)]
+          // make configurable
         });
         result.text = text;
         inputTokens = totalUsage?.inputTokens || 0;
@@ -10684,6 +10862,7 @@ var providerRateLimiter = async (key, windowSeconds, limit, points) => {
 };
 
 // src/exulu/routes.ts
+var import_zod_from_json_schema = require("zod-from-json-schema");
 var REQUEST_SIZE_LIMIT = "50mb";
 var getExuluVersionNumber = async () => {
   try {
@@ -11104,6 +11283,16 @@ Mood: friendly and intelligent.
         providers,
         user
       );
+      if (req.body.outputSchema && !!headers.stream) {
+        throw new Error("Providing a outputSchema in the POST body is not allowed when using the streaming API, set 'stream' to false in the headers when defining a response schema.");
+      }
+      let outputSchema;
+      if (req.body.outputSchema) {
+        if (typeof req.body.outputSchema === "string") {
+          req.body.outputSchema = JSON.parse(req.body.outputSchema);
+        }
+        outputSchema = (0, import_zod_from_json_schema.convertJsonSchemaToZod)(req.body.outputSchema);
+      }
       let providerapikey;
       const variableName = agent.providerapikey;
       if (variableName) {
@@ -11250,6 +11439,7 @@ Mood: friendly and intelligent.
         const response = await provider.generateSync({
           contexts,
           rerankers: rerankers || [],
+          outputSchema,
           agent,
           user,
           req,
@@ -15485,6 +15675,22 @@ var MarkdownChunker = class {
     });
     return result;
   }
+  /**
+   * Checks if a position in the text falls within a <diagram> tag.
+   * Returns the adjusted position (before the diagram) if inside a diagram, otherwise returns the original position.
+   */
+  adjustForDiagramTags(text, position) {
+    const diagramRegex = /<diagram>[\s\S]*?<\/diagram>/gi;
+    let match;
+    while ((match = diagramRegex.exec(text)) !== null) {
+      const diagramStart = match.index;
+      const diagramEnd = match.index + match[0].length;
+      if (position > diagramStart && position < diagramEnd) {
+        return diagramStart;
+      }
+    }
+    return position;
+  }
   /**
    * Find the nearest logical breakpoint working backwards from the end of the text.
    * Logical breakpoints are prioritized as follows:
@@ -15496,6 +15702,7 @@ var MarkdownChunker = class {
    *
    * Only considers breakpoints in the last 50% of the text to avoid creating very small chunks.
    * Returns the position of the breakpoint, or null if none found
+   * IMPORTANT: Never splits content within <diagram> tags
    */
   findLogicalBreakpoint(text) {
     if (text.length === 0) return null;
@@ -15515,7 +15722,7 @@ var MarkdownChunker = class {
       }
     }
     if (lastHeaderPosition > 0) {
-      return lastHeaderPosition;
+      return this.adjustForDiagramTags(text, lastHeaderPosition);
     }
     let lastParagraphBreak = -1;
     let searchPos = text.length;
@@ -15528,11 +15735,12 @@ var MarkdownChunker = class {
       searchPos = pos;
     }
     if (lastParagraphBreak > 0) {
-      return lastParagraphBreak + 2;
+      const adjusted = this.adjustForDiagramTags(text, lastParagraphBreak + 2);
+      return adjusted;
     }
     const newlineIndex = text.lastIndexOf("\n");
     if (newlineIndex >= minPosition) {
-      return newlineIndex + 1;
+      return this.adjustForDiagramTags(text, newlineIndex + 1);
     }
     const sentenceEndRegex = /[.!?](?:\s|$)/g;
     let lastSentenceEnd = -1;
@@ -15542,13 +15750,13 @@ var MarkdownChunker = class {
       }
     }
     if (lastSentenceEnd > 0) {
-      return lastSentenceEnd;
+      return this.adjustForDiagramTags(text, lastSentenceEnd);
     }
     let lastSpace = text.length;
     while (lastSpace > minPosition) {
       const pos = text.lastIndexOf(" ", lastSpace - 1);
       if (pos >= minPosition) {
-        return pos + 1;
+        return this.adjustForDiagramTags(text, pos + 1);
       }
       lastSpace = pos;
     }
@@ -15680,6 +15888,38 @@ var MarkdownChunker = class {
           targetPosition = currentPosition + decoded.length;
         }
       }
+      const diagramCheck = /<diagram>/gi;
+      const diagramCloseCheck = /<\/diagram>/gi;
+      let openDiagramsInSlice = 0;
+      while (diagramCheck.exec(currentSlice) !== null) {
+        openDiagramsInSlice++;
+      }
+      let closeDiagramsInSlice = 0;
+      while (diagramCloseCheck.exec(currentSlice) !== null) {
+        closeDiagramsInSlice++;
+      }
+      if (openDiagramsInSlice > closeDiagramsInSlice) {
+        const lastDiagramOpenIndex = currentSlice.lastIndexOf("<diagram>");
+        if (lastDiagramOpenIndex !== -1) {
+          const remainingText = text.slice(currentPosition + lastDiagramOpenIndex);
+          const closingTagMatch = /<\/diagram>/i.exec(remainingText);
+          if (closingTagMatch) {
+            const closingTagPosition = lastDiagramOpenIndex + closingTagMatch.index + closingTagMatch[0].length;
+            const extendedSlice = text.slice(currentPosition, currentPosition + closingTagPosition);
+            const extendedTokens = tokenizer.encode(extendedSlice);
+            if (extendedTokens.length <= adjustedChunkSize * 1.5) {
+              currentSlice = extendedSlice;
+              targetPosition = currentPosition + closingTagPosition;
+            } else {
+              currentSlice = currentSlice.slice(0, lastDiagramOpenIndex);
+              targetPosition = currentPosition + lastDiagramOpenIndex;
+            }
+          } else {
+            currentSlice = currentSlice.slice(0, lastDiagramOpenIndex);
+            targetPosition = currentPosition + lastDiagramOpenIndex;
+          }
+        }
+      }
       const breakpointPosition = this.findLogicalBreakpoint(currentSlice);
       if (breakpointPosition !== null) {
         currentSlice = currentSlice.slice(0, breakpointPosition);
@@ -16101,6 +16341,8 @@ ${command}`;
 }
 
 // ee/python/documents/processing/doc_processor.ts
+var import_liteparse = require("@llamaindex/liteparse");
+var import_mistralai = require("@mistralai/mistralai");
 async function processDocx(file) {
   const html = await mammoth.convertToHtml({ buffer: file });
   const turndownService = new import_turndown.default();
@@ -16175,43 +16417,80 @@ async function validatePageWithVLM(page, imagePath, model) {
   const imageBuffer = await fs2.promises.readFile(imagePath);
   const imageBase64 = imageBuffer.toString("base64");
   const mimeType = "image/png";
-  const prompt = `You are validating OCR/document parsing output for a page that might contain tables and images.
-
-Here is the current OCR/parsed content for this page:
+  const prompt = `You are a document validation assistant. Your task is to analyze a page image and correct the output of an OCR/parsing pipeline. The content may include tables, technical diagrams, schematics, and structured text.
 
 ---
+## CURRENT OCR OUTPUT
+
 ${page.content}
 ---
 
-Please analyze the page image and validate it:
-
-1. Check if the extracted markdown text accurately represents the content from the page, including:
-   - Table data (rows, columns, headers, values)
-   - Technical diagrams, schematics, control boards
-   - Icons, checkmarks, symbols
-   - Image captions and labels
-
-2. If the page has significant errors or omissions, provide a corrected version for the page.
+## YOUR TASK
 
-3. Return a validation result for the page.
+Compare the page image to the OCR output above. Identify errors, omissions, and formatting issues, then return a structured validation result (see OUTPUT FORMAT below).
 
-IMPORTANT OUTPUT FORMAT REQUIREMENTS:
-- You MUST output all tables in proper Markdown table format using pipes (|) and dashes (---)
-- Use simple separator rows: | --- | --- | (NOT long dashes like ----------------------)
-- Every table must have: header row, separator row, and data rows
-- Example format:
+---
+## VALIDATION CHECKLIST
+
+Work through these checks in order:
+
+### 1. Text Accuracy
+- Verify all text is correctly transcribed.
+- For minor character-level OCR errors (e.g. "\xF6" vs "\xFC", "rn" vs "m"), **prefer the original OCR output** unless you are certain of an error. Do not silently "fix" characters based on guesswork.
+
+### 2. Heading Levels
+- Verify that headings use correct Markdown levels (#, ##, ###, ####, #####, ######).
+- Determine heading level using the following priority:
+  1. **Hierarchical numbering** (strongest signal): e.g. "1" \u2192 #, "2.1" \u2192 ##, "2.1.1" \u2192 ###, "2.1.2.5" \u2192 ####
+  2. Font size (larger = higher level)
+  3. Indentation
+  4. Bold/emphasis styling
+
+### 3. Tables
+
+**First, decide whether the table should be Markdown or plain text:**
+- Use **Markdown table format** if the table has a consistent, clear header structure and uniform column layout throughout.
+- Use **plain text structured description** if the table:
+  - Lacks a clear header row
+  - Uses mixed or irregular column structures across rows
+  - Functions more like a certificate, form, or label layout
+
+**If using Markdown format**, follow these rules strictly:
+- Every table must have: header row \u2192 separator row \u2192 data rows
+- Use simple separators only: \`| --- | --- |\` (NOT \`|---|---|\` or long dashes)
+- Example:
+  \`\`\`
   | Column 1 | Column 2 |
   | --- | --- |
-  | Data 1 | Data 2 |
-- If the extracted content already has tables, preserve their structure but fix any errors you find in the actual data
-- Do NOT output tables as plain text or in any other format
-- Preserve all markdown formatting (headings with ##, lists, etc.)
-
-Specific notes and guidelines:
-- Some pages might contain a table with a column that show black and white dots (for Example Rufe-LEDs). You should translate this into + for black (meaning active) and - for white (meaning inactive).
-- Some tables might use green or black checkmarks and red or black crosses. You should translate this into + for checkmarks (meaning active) and - for a cross (meaning inactive).
-- IMPORTANT: Only provide corrections if you find actual errors in the content. If the extracted text is accurate, set needs_correction to false.
-
+  | Data 1   | Data 2   |
+  \`\`\`
+- Important: do not use the | character as part of the data inside a cell, this would break the table, if a cell contains a | character, use a capital I.
+
+**Symbol translation rules for table cells:**
+- Black/filled dot \u2192 \`+\` (active); White/empty dot \u2192 \`-\` (inactive)  
+  *(e.g. Rufe-LED columns)*
+- Green or black checkmark \u2192 \`+\` (active); Red or black cross \u2192 \`-\` (inactive)
+
+### 4. Multi-Page Table Continuity
+- If this page contains a table with a header row that runs to the bottom of the page (suggesting it may continue on the next page), extract the header row and include it in the \`current_page_table.headers\` field.
+- If this page contains a table WITHOUT a header row (suggesting it's a continuation from a previous page), set \`current_page_table.is_continuation\` to true and try to identify what the headers might be based on the data structure. Include your best guess for headers in \`current_page_table.headers\`.
+
+### 5. Technical Diagrams & Schematics
+If the page contains a flow-chart, schematic, technical drawing or control board layout that is **absent or poorly described** in the OCR output do the following:
+- Open a <diagram> tag with the following content:
+  <diagram>
+    <description>
+      Add a detailed description of the diagram here.
+    </description>
+    <mermaid>
+      Add a mermaid diagram schema here that in detail describes the diagram.
+    </mermaid>
+  </diagram>
+
+### 6. Captions, Icons & Symbols
+- Verify that image captions, labels, icons, and checkmarks are present and correctly transcribed.
+
+### 7. Only populate \`corrected_text\` when \`needs_correction\` is true. If the OCR output is accurate, return \`needs_correction: false\` and \`corrected_content: null\`.
 `;
   const result = await (0, import_ai7.generateText)({
     model,
@@ -16219,6 +16498,10 @@ Specific notes and guidelines:
       schema: import_zod12.z.object({
         needs_correction: import_zod12.z.boolean(),
         corrected_text: import_zod12.z.string().nullable(),
+        current_page_table: import_zod12.z.object({
+          headers: import_zod12.z.array(import_zod12.z.string()),
+          is_continuation: import_zod12.z.boolean()
+        }).nullable(),
         confidence: import_zod12.z.enum(["high", "medium", "low"]),
         reasoning: import_zod12.z.string()
       })
@@ -16241,23 +16524,80 @@ Specific notes and guidelines:
     needs_correction: parsedOutput.needs_correction,
     corrected_text: parsedOutput.corrected_text || void 0,
     confidence: parsedOutput.confidence,
+    current_page_table: parsedOutput.current_page_table || void 0,
     reasoning: parsedOutput.reasoning
   };
   return validation;
 }
+function reconstructTableHeaders(document2, validationResults, verbose = false) {
+  let lastTableHeaders = void 0;
+  for (const page of document2) {
+    const validation = validationResults.get(page.page);
+    if (!validation) continue;
+    const tableInfo = validation.current_page_table;
+    if (tableInfo && tableInfo.headers.length > 0) {
+      if (tableInfo.is_continuation && lastTableHeaders) {
+        if (verbose) {
+          console.log(`[EXULU] Page ${page.page}: Reconstructing table headers from previous page`);
+          console.log(`[EXULU] Previous headers: ${lastTableHeaders.join(" | ")}`);
+        }
+        const contentToModify = page.vlm_corrected_text || page.content;
+        const lines = contentToModify.split("\n");
+        const firstTableLineIndex = lines.findIndex((line) => line.trim().startsWith("|"));
+        if (firstTableLineIndex !== -1) {
+          const headerRow = `| ${lastTableHeaders.join(" | ")} |`;
+          const separatorRow = `| ${lastTableHeaders.map(() => "---").join(" | ")} |`;
+          lines.splice(firstTableLineIndex, 0, headerRow, separatorRow);
+          const reconstructedContent = lines.join("\n");
+          if (page.vlm_corrected_text) {
+            page.vlm_corrected_text = reconstructedContent;
+          } else {
+            page.content = reconstructedContent;
+          }
+          if (verbose) {
+            console.log(`[EXULU] Page ${page.page}: Added table headers successfully`);
+          }
+        }
+        if (!tableInfo.is_continuation) {
+          lastTableHeaders = tableInfo.headers;
+        }
+      } else {
+        lastTableHeaders = tableInfo.headers;
+        if (verbose) {
+          console.log(`[EXULU] Page ${page.page}: Storing table headers for potential continuation`);
+          console.log(`[EXULU] Headers: ${lastTableHeaders.join(" | ")}`);
+        }
+      }
+    } else {
+      lastTableHeaders = void 0;
+    }
+  }
+}
 async function validateWithVLM(document2, model, verbose = false, concurrency = 10) {
   console.log(`[EXULU] Starting VLM validation for docling output, ${document2.length} pages...`);
-  console.log(
-    `[EXULU] Concurrency limit: ${concurrency}`
-  );
+  console.log(`[EXULU] Concurrency limit: ${concurrency}`);
+  const limit = (0, import_p_limit.default)(concurrency);
+  const validationResults = /* @__PURE__ */ new Map();
   let validatedCount = 0;
   let correctedCount = 0;
-  const limit = (0, import_p_limit.default)(concurrency);
   const validationTasks = document2.map(
     (page) => limit(async () => {
+      await new Promise((resolve3) => setImmediate(resolve3));
       const imagePath = page.image;
+      if (!page.content) {
+        console.warn(`[EXULU] Page ${page.page}: No content found, skipping validation`);
+        return;
+      }
       if (!imagePath) {
-        console.log(`[EXULU] Page ${page.page}: No image found, skipping validation`);
+        console.warn(`[EXULU] Page ${page.page}: No image found, skipping validation`);
+        return;
+      }
+      const hasImage = page.content.match(/\.(jpeg|jpg|png|gif|webp)/i);
+      const hasTable = (page.content.match(/\|/g)?.length || 0) > 1;
+      if (!hasImage && !hasTable) {
+        if (verbose) {
+          console.log(`[EXULU] Page ${page.page}: No image or table found, SKIPPING VLM validation`);
+        }
         return;
       }
       let validation;
@@ -16265,6 +16605,13 @@ async function validateWithVLM(document2, model, verbose = false, concurrency =
         validation = await withRetry(async () => {
           return await validatePageWithVLM(page, imagePath, model);
         }, 3);
+        validationResults.set(page.page, validation);
+        if (verbose && validation.current_page_table) {
+          console.log(`[EXULU] Page ${page.page} table info:`, {
+            headers: validation.current_page_table.headers,
+            is_continuation: validation.current_page_table.is_continuation
+          });
+        }
       } catch (error) {
         console.error(`[EXULU] Error validating page ${page.page} with VLM more than 3 times, skipping:`, error);
         throw error;
@@ -16295,9 +16642,12 @@ async function validateWithVLM(document2, model, verbose = false, concurrency =
     })
   );
   await Promise.all(validationTasks);
-  console.log(`[EXULU] VLM validation complete:`);
-  console.log(`[EXULU] Validated: ${validatedCount} chunks`);
-  console.log(`[EXULU] Corrected: ${correctedCount} chunks`);
+  console.log(`[EXULU] VLM validation complete (parallel processing):`);
+  console.log(`[EXULU] Validated: ${validatedCount} pages`);
+  console.log(`[EXULU] Corrected: ${correctedCount} pages`);
+  console.log(`[EXULU] Starting sequential table header reconstruction...`);
+  reconstructTableHeaders(document2, validationResults, verbose);
+  console.log(`[EXULU] Table header reconstruction complete`);
   return document2;
 }
 async function processDocument(filePath, fileType, buffer, tempDir, config, verbose = false) {
@@ -16312,15 +16662,6 @@ async function processDocument(filePath, fileType, buffer, tempDir, config, verb
   const stripped = filePath.split(".").pop()?.trim();
   let result;
   switch (stripped) {
-    case "pdf":
-      result = await processPdf(buffer, paths, config, verbose);
-      break;
-    case "docx":
-      result = await processDocx(buffer);
-      break;
-    case "doc":
-      result = await processWord(buffer);
-      break;
     case "txt":
     case "md":
       let content = buffer.toString();
@@ -16334,6 +16675,15 @@ async function processDocument(filePath, fileType, buffer, tempDir, config, verb
         }]
       };
       break;
+    case "pdf":
+      result = await processPdf(buffer, paths, config, verbose);
+      break;
+    case "docx":
+      result = await processDocx(buffer);
+      break;
+    case "doc":
+      result = await processWord(buffer);
+      break;
     // Todo other file types with docx and officeparser
     default:
       throw new Error(`[EXULU] Unsupported file type: ${fileType}`);
@@ -16348,8 +16698,8 @@ async function processDocument(filePath, fileType, buffer, tempDir, config, verb
 }
 async function processPdf(buffer, paths, config, verbose = false) {
   try {
-    let json;
-    if (config?.docling) {
+    let json = [];
+    if (config?.processor.name === "docling") {
       console.log(`[EXULU] Validating Python environment...`);
       const validation = await validatePythonEnvironment(void 0, true);
       if (!validation.valid) {
@@ -16390,7 +16740,7 @@ ${setupResult.output || ""}`);
       }
       const jsonContent = await fs2.promises.readFile(paths.json, "utf-8");
       json = JSON.parse(jsonContent);
-    } else {
+    } else if (config?.processor.name === "officeparser") {
       const text = await (0, import_officeparser2.parseOfficeAsync)(buffer, {
         outputErrorToConsole: false,
         newlineDelimiter: "\n"
@@ -16400,15 +16750,69 @@ ${setupResult.output || ""}`);
         content: text,
         headings: []
       }];
+    } else if (config?.processor.name === "mistral") {
+      if (!process.env.MISTRAL_API_KEY) {
+        throw new Error("[EXULU] MISTRAL_API_KEY is not set, please set it in the environment variables.");
+      }
+      await new Promise((resolve3) => setTimeout(resolve3, Math.floor(Math.random() * 4e3) + 1e3));
+      const base64Pdf = buffer.toString("base64");
+      const client2 = new import_mistralai.Mistral({ apiKey: process.env.MISTRAL_API_KEY });
+      const ocrResponse = await withRetry(async () => {
+        const ocrResponse2 = await client2.ocr.process({
+          document: {
+            type: "document_url",
+            documentUrl: "data:application/pdf;base64," + base64Pdf
+          },
+          model: "mistral-ocr-latest",
+          includeImageBase64: false
+        });
+        return ocrResponse2;
+      }, 10);
+      const parser = new import_liteparse.LiteParse();
+      const screenshots = await parser.screenshot(paths.source, void 0);
+      await fs2.promises.mkdir(paths.images, { recursive: true });
+      for (const screenshot of screenshots) {
+        await fs2.promises.writeFile(
+          path.join(
+            paths.images,
+            `${screenshot.pageNum}.png`
+          ),
+          screenshot.imageBuffer
+        );
+        screenshot.imagePath = path.join(paths.images, `${screenshot.pageNum}.png`);
+      }
+      json = ocrResponse.pages.map((page) => ({
+        page: page.index + 1,
+        content: page.markdown,
+        image: screenshots.find((s) => s.pageNum === page.index + 1)?.imagePath,
+        headings: []
+      }));
+      fs2.writeFileSync(paths.json, JSON.stringify(json, null, 2));
+    } else if (config?.processor.name === "liteparse") {
+      const parser = new import_liteparse.LiteParse();
+      const result = await parser.parse(paths.source);
+      const screenshots = await parser.screenshot(paths.source, void 0);
+      console.log(`[EXULU] Liteparse screenshots: ${JSON.stringify(screenshots)}`);
+      await fs2.promises.mkdir(paths.images, { recursive: true });
+      for (const screenshot of screenshots) {
+        await fs2.promises.writeFile(path.join(paths.images, `${screenshot.pageNum}.png`), screenshot.imageBuffer);
+        screenshot.imagePath = path.join(paths.images, `${screenshot.pageNum}.png`);
+      }
+      json = result.pages.map((page) => ({
+        page: page.pageNum,
+        content: page.text,
+        image: screenshots.find((s) => s.pageNum === page.pageNum)?.imagePath
+      }));
+      fs2.writeFileSync(paths.json, JSON.stringify(json, null, 2));
     }
     console.log(`[EXULU] 
 \u2713 Document processing completed successfully`);
     console.log(`[EXULU] Total pages: ${json.length}`);
     console.log(`[EXULU] Output file: ${paths.json}`);
-    if (!config?.docling && config?.vlm?.model) {
+    if (config?.vlm?.model) {
       console.error("[EXULU] VLM validation is only supported when docling is enabled, skipping validation.");
     }
-    if (config?.docling && config?.vlm?.model) {
+    if (config?.vlm?.model && json.length > 0) {
       json = await validateWithVLM(
         json,
         config.vlm.model,
@@ -16436,29 +16840,37 @@ ${setupResult.output || ""}`);
         "utf-8"
       );
     }
-    const markdown = json.map((p) => {
-      if (p.vlm_corrected_text) {
-        return p.vlm_corrected_text;
-      } else {
-        return p.content;
+    const markdownStream = fs2.createWriteStream(paths.markdown, { encoding: "utf-8" });
+    for (let i = 0; i < json.length; i++) {
+      const p = json[i];
+      if (!p) continue;
+      const content = p.vlm_corrected_text ?? p.content;
+      markdownStream.write(content);
+      if (i < json.length - 1) {
+        markdownStream.write("\n\n\n<!-- END_OF_PAGE -->\n\n\n");
       }
-    }).join("\n\n\n<!-- END_OF_PAGE -->\n\n\n");
-    await fs2.promises.writeFile(
-      paths.markdown,
-      markdown,
-      "utf-8"
-    );
+    }
+    await new Promise((resolve3, reject) => {
+      markdownStream.end(() => resolve3());
+      markdownStream.on("error", reject);
+    });
     console.log(`[EXULU] Validated output saved to: ${paths.json}`);
     console.log(`[EXULU] Validated markdown saved to: ${paths.markdown}`);
+    const markdown = await fs2.promises.readFile(paths.markdown, "utf-8");
+    const processedJson = json.map((e) => {
+      const finalContent = e.vlm_corrected_text ?? e.content;
+      return {
+        page: e.page,
+        content: finalContent
+      };
+    });
+    json.length = 0;
+    json = [];
+    const memUsage = process.memoryUsage();
+    console.log(`[EXULU] Memory after document processing: ${Math.round(memUsage.heapUsed / 1024 / 1024)}MB / ${Math.round(memUsage.heapTotal / 1024 / 1024)}MB`);
     return {
       markdown,
-      json: json.map((e) => {
-        const finalContent = e.vlm_corrected_text || e.content;
-        return {
-          page: e.page,
-          content: finalContent
-        };
-      })
+      json: processedJson
     };
   } catch (error) {
     console.error("[EXULU] Error processing document:", error);
@@ -16471,9 +16883,9 @@ var loadFile = async (file, name, tempDir) => {
   if (!fileType) {
     throw new Error("[EXULU] File name does not include extension, extension is required for document processing.");
   }
+  const UUID = (0, import_crypto.randomUUID)();
   let buffer;
   if (Buffer.isBuffer(file)) {
-    const UUID = (0, import_crypto.randomUUID)();
     filePath = path.join(tempDir, `${UUID}.${fileType}`);
     await fs2.promises.writeFile(filePath, file);
     buffer = file;
@@ -16482,7 +16894,10 @@ var loadFile = async (file, name, tempDir) => {
     if (filePath.startsWith("http")) {
       const response = await fetch(filePath);
       const array = await response.arrayBuffer();
+      const tempFilePath = path.join(tempDir, `${UUID}.${fileType}`);
+      await fs2.promises.writeFile(tempFilePath, Buffer.from(array));
       buffer = Buffer.from(array);
+      filePath = tempFilePath;
     } else {
       buffer = await fs2.promises.readFile(file);
     }
@@ -16500,17 +16915,34 @@ async function documentProcessor({
   }
   const uuid = (0, import_crypto.randomUUID)();
   const tempDir = path.join(process.cwd(), "temp", uuid);
+  const localFilesAndFoldersToDelete = [tempDir];
   console.log(`[EXULU] Temporary directory for processing document ${name}: ${tempDir}`);
   await fs2.promises.mkdir(tempDir, { recursive: true });
+  const timestamp = (/* @__PURE__ */ new Date()).toISOString();
+  await fs2.promises.writeFile(path.join(tempDir, "created_at.txt"), timestamp);
   try {
     const {
       filePath,
       fileType,
       buffer
     } = await loadFile(file, name, tempDir);
-    const supportedTypes = ["pdf", "docx", "doc", "txt", "md"];
+    let supportedTypes = [];
+    switch (config?.processor.name) {
+      case "docling":
+        supportedTypes = ["pdf", "docx", "doc", "txt", "md"];
+        break;
+      case "officeparser":
+        supportedTypes = [];
+        break;
+      case "liteparse":
+        supportedTypes = ["pdf", "doc", "docx", "docm", "odt", "rtf", "ppt", "pptx", "pptm", "odp", "xls", "xlsx", "xlsm", "ods", "csv", "tsv"];
+        break;
+      case "mistral":
+        supportedTypes = ["pdf", "docx", "doc", "txt", "md"];
+        break;
+    }
     if (!supportedTypes.includes(fileType)) {
-      throw new Error(`[EXULU] Unsupported file type: ${fileType} for Exulu document processor.`);
+      throw new Error(`[EXULU] Unsupported file type: ${fileType} for Exulu document processor, the ${config?.processor.name} processor only supports the following file types: ${supportedTypes.join(", ")}.`);
     }
     const { content } = await processDocument(
       filePath,
@@ -16523,9 +16955,19 @@ async function documentProcessor({
     return content.json;
   } catch (error) {
     console.error("Error during chunking:", error);
-    return void 0;
+    throw error;
   } finally {
-    await fs2.promises.rm(tempDir, { recursive: true });
+    if (config?.debugging?.deleteTempFiles !== false) {
+      for (const file2 of localFilesAndFoldersToDelete) {
+        try {
+          await fs2.promises.rm(file2, { recursive: true });
+          console.log(`[EXULU] Deleted file or folder: ${file2}`);
+        } catch (error) {
+          console.error(`[EXULU] Error deleting file or folder: ${file2}`, error);
+          console.log(`[EXULU] File or folder still exists: ${file2}`);
+        }
+      }
+    }
   }
 }