@exulu/backend 3.7.1 → 3.7.3

package/dist/index.cjs

@@ -10212,17 +10212,31 @@ function detectQueryLanguage(query, minLength = 10) {
 }
 function stemWord(word, languageCode) {
   const stemmer = STEMMER_MAP[languageCode] || import_natural.default.PorterStemmer;
-  const cleaned = word.replace(/[^\p{L}\p{N}]/gu, "").toLowerCase();
-  if (!cleaned) {
+  const core = word.replace(/^[^\p{L}\p{N}]+/u, "").replace(/[^\p{L}\p{N}]+$/u, "").toLowerCase();
+  if (!core) {
     return word;
   }
+  if (!/^\p{L}+$/u.test(core)) {
+    return core;
+  }
   try {
-    return stemmer.stem(cleaned);
+    return stemmer.stem(core);
   } catch (error) {
     console.warn(`[EXULU] Error stemming word "${word}":`, error);
-    return cleaned;
+    return core;
   }
 }
+function buildFullTextOrQuery(original, processed) {
+  const seen = /* @__PURE__ */ new Set();
+  const tokens = [];
+  for (const raw of `${original} ${processed}`.split(/\s+/)) {
+    const token = raw.replace(/["'`]/g, "").replace(/^[^\p{L}\p{N}]+/u, "").replace(/[^\p{L}\p{N}]+$/u, "").toLowerCase();
+    if (!token || token === "or" || seen.has(token)) continue;
+    seen.add(token);
+    tokens.push(token);
+  }
+  return tokens.join(" or ");
+}
 function preprocessQuery(query, options = {}) {
   const {
     enableStemming = true,
@@ -10257,6 +10271,18 @@ function preprocessQuery(query, options = {}) {
     stemmed: true
   };
 }
+function resolveSearchQueryTexts(query) {
+  const { processed } = preprocessQuery(query, {
+    enableStemming: true,
+    detectLanguage: true
+  });
+  const ftsText = processed || query;
+  return {
+    embedText: query,
+    ftsText,
+    hybridOrQuery: buildFullTextOrQuery(query, ftsText)
+  };
+}
 
 // src/graphql/resolvers/apply-sorting.ts
 init_cjs_shims();
@@ -12735,17 +12761,15 @@ var vectorSearch = async ({
   const _tBody = Date.now();
   let _preMs = 0, _statMs = 0, _resolveMs = 0, _embedMs = 0;
   let _embedSource = "none";
+  let hybridOrQuery = "";
   if (query) {
     const _tp = Date.now();
-    const { processed: stemmedQuery } = preprocessQuery(query, {
-      enableStemming: true,
-      detectLanguage: true
-    });
+    const texts = resolveSearchQueryTexts(query);
     _preMs = Date.now() - _tp;
-    console.log("[EXULU] Stemmed query:", stemmedQuery);
-    if (stemmedQuery) {
-      query = stemmedQuery;
-    }
+    console.log("[EXULU] Search query texts:", texts);
+    const embedText = texts.embedText;
+    hybridOrQuery = texts.hybridOrQuery;
+    query = texts.ftsText;
     if (queryEmbedding && queryEmbedding.length) {
       vector = queryEmbedding;
       _embedSource = "reused";
@@ -12771,7 +12795,7 @@ var vectorSearch = async ({
       });
       _resolveMs = Date.now() - _tr;
       const _te = Date.now();
-      const [queryVector] = await resolved.embed([query], { inputType: "query" });
+      const [queryVector] = await resolved.embed([embedText], { inputType: "query" });
       _embedMs = Date.now() - _te;
       if (!queryVector?.length) {
         throw new Error("No vector generated for query.");
@@ -12843,10 +12867,10 @@ var vectorSearch = async ({
       const fullTextWeight = 2;
       const semanticWeight = 1;
       const rrfK = 50;
-      const ftRankExpression = languages.map((lang) => `ts_rank(chunks.fts, plainto_tsquery('${lang}', ?))`).join(", ");
-      const ftRankParams = languages.map(() => query);
-      const ftMatchExpression = languages.map((lang) => `chunks.fts @@ plainto_tsquery('${lang}', ?)`).join(" OR ");
-      const ftMatchParams = languages.map(() => query);
+      const ftRankExpression = languages.map((lang) => `ts_rank(chunks.fts, websearch_to_tsquery('${lang}', ?))`).join(", ");
+      const ftRankParams = languages.map(() => hybridOrQuery);
+      const ftMatchExpression = languages.map((lang) => `chunks.fts @@ websearch_to_tsquery('${lang}', ?)`).join(" OR ");
+      const ftMatchParams = languages.map(() => hybridOrQuery);
       let fullTextQuery = db2(chunksTable + " as chunks").select([
         "chunks.id",
         "chunks.source",
@@ -12886,8 +12910,8 @@ var vectorSearch = async ({
         db2.raw('items."updatedAt" as item_updated_at'),
         db2.raw('items."createdAt" as item_created_at'),
         db2.raw(
-          `GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, plainto_tsquery('${lang}', ?))`).join(", ")}) AS fts_rank`,
-          languages.map(() => query)
+          `GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, websearch_to_tsquery('${lang}', ?))`).join(", ")}) AS fts_rank`,
+          languages.map(() => hybridOrQuery)
         ),
         db2.raw(`(1 - (chunks.embedding <=> ${vectorExpr})) AS cosine_distance`),
         db2.raw(
@@ -12912,8 +12936,8 @@ var vectorSearch = async ({
                   `,
         [rrfK, fullTextWeight, rrfK, semanticWeight, cutoffs?.hybrid || 0]
       ).whereRaw(
-        `(chunks.fts IS NULL OR GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, plainto_tsquery('${lang}', ?))`).join(", ")}) > ?)`,
-        [...languages.map(() => query), cutoffs?.tsvector || 0]
+        `(chunks.fts IS NULL OR GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, websearch_to_tsquery('${lang}', ?))`).join(", ")}) > ?)`,
+        [...languages.map(() => hybridOrQuery), cutoffs?.tsvector || 0]
       ).whereRaw(`(chunks.embedding IS NULL OR (1 - (chunks.embedding <=> ${vectorExpr})) >= ?)`, [
         cutoffs?.cosineDistance || 0
       ]).orderByRaw("hybrid_score DESC").limit(Math.min(matchCount, 250));
@@ -16658,6 +16682,37 @@ var bullmq = {
   }
 };
 
+// src/utils/serialize-error.ts
+init_cjs_shims();
+var MAX_CAUSE_DEPTH = 3;
+function serializeError(err, depth = 0) {
+  if (err instanceof Error) {
+    const out = { name: err.name, message: err.message };
+    if (err.stack) out.stack = err.stack;
+    for (const [key, value] of Object.entries(err)) {
+      if (key !== "cause") out[key] = value;
+    }
+    const cause = err.cause;
+    if (cause !== void 0 && depth < MAX_CAUSE_DEPTH) {
+      out.cause = serializeError(cause, depth + 1);
+    }
+    return out;
+  }
+  if (typeof err === "string") {
+    return { message: err };
+  }
+  if (err !== null && typeof err === "object") {
+    let message;
+    try {
+      message = JSON.stringify(err);
+    } catch {
+      message = "[unserializable error object]";
+    }
+    return { ...err, message };
+  }
+  return { message: String(err) };
+}
+
 // src/utils/enabled-tools.ts
 init_cjs_shims();
 init_pipeline();
@@ -16676,6 +16731,114 @@ init_statistics();
 
 // src/exulu/generate-stream.ts
 init_cjs_shims();
+
+// src/exulu/file-part.ts
+init_cjs_shims();
+var IMAGE_EXTENSION_MIME = {
+  ".png": "image/png",
+  ".jpeg": "image/jpeg",
+  ".jpg": "image/jpeg",
+  ".gif": "image/gif",
+  ".webp": "image/webp",
+  // Windows bitmaps arrive as email logos / signature images and camera exports.
+  // Left out, they were routed to OfficeParser and replaced by an error text.
+  ".bmp": "image/bmp",
+  ".dib": "image/bmp"
+};
+var MEDIA_TYPE_ALIASES = {
+  "image/jpg": "image/jpeg",
+  "image/x-ms-bmp": "image/bmp",
+  "image/x-bmp": "image/bmp",
+  "image/x-windows-bmp": "image/bmp"
+};
+function fileNameFromUrl(url) {
+  if (!url) return void 0;
+  let path4;
+  try {
+    path4 = new URL(url).pathname;
+  } catch {
+    path4 = url.split("?")[0] ?? "";
+  }
+  const last = path4.split("/").pop() ?? "";
+  try {
+    return decodeURIComponent(last) || void 0;
+  } catch {
+    return last || void 0;
+  }
+}
+function classifyFilePart(part) {
+  const filename = part.filename ?? fileNameFromUrl(part.url);
+  const lowerName = (filename ?? "").toLowerCase();
+  const extension = Object.keys(IMAGE_EXTENSION_MIME).find((ext) => lowerName.endsWith(ext));
+  const declared = typeof part.mediaType === "string" ? part.mediaType.trim().toLowerCase() : "";
+  const declaredIsImage = declared.startsWith("image/");
+  if (declaredIsImage || extension) {
+    const mediaType = declaredIsImage ? MEDIA_TYPE_ALIASES[declared] ?? declared : IMAGE_EXTENSION_MIME[extension];
+    return { kind: "image", mediaType, filename };
+  }
+  return { kind: "document", filename: filename ?? "attachment" };
+}
+
+// src/exulu/stored-messages.ts
+init_cjs_shims();
+function dropEmptyMessages(messages) {
+  const kept = messages.filter((m) => Array.isArray(m.parts) && m.parts.length > 0);
+  return kept.length === messages.length ? messages : kept;
+}
+
+// src/exulu/stored-file-url.ts
+init_cjs_shims();
+var stripTrailingSlash = (s) => s.replace(/\/+$/, "");
+function decodeKey(rawPath) {
+  const key = rawPath.replace(/^\/+/, "");
+  try {
+    return decodeURIComponent(key);
+  } catch {
+    return key;
+  }
+}
+function parseStoredFileUrl(url, store) {
+  if (!url || !store.bucket) return void 0;
+  let parsed;
+  try {
+    parsed = new URL(url);
+  } catch {
+    return void 0;
+  }
+  if (parsed.protocol !== "http:" && parsed.protocol !== "https:") return void 0;
+  if (store.endpoint) {
+    let endpoint;
+    try {
+      endpoint = new URL(store.endpoint);
+    } catch {
+      return void 0;
+    }
+    if (parsed.host !== endpoint.host) return void 0;
+    const base = stripTrailingSlash(endpoint.pathname);
+    const prefix = `${base}/${store.bucket}/`;
+    if (!parsed.pathname.startsWith(prefix)) return void 0;
+    const key2 = decodeKey(parsed.pathname.slice(prefix.length));
+    return key2 ? { bucket: store.bucket, key: key2 } : void 0;
+  }
+  if (!parsed.host.startsWith(`${store.bucket}.s3`) || !parsed.host.endsWith(".amazonaws.com")) {
+    return void 0;
+  }
+  const key = decodeKey(parsed.pathname);
+  return key ? { bucket: store.bucket, key } : void 0;
+}
+async function resolveFreshFileUrl(url, opts) {
+  const location = parseStoredFileUrl(url, opts);
+  if (!location) return url;
+  try {
+    return await opts.sign(location.bucket, location.key);
+  } catch (err) {
+    console.warn(`[EXULU] could not re-sign stored file URL for key "${location.key}":`, err);
+    return url;
+  }
+}
+
+// src/exulu/generate-stream.ts
+init_uppy();
 var import_ai6 = require("ai");
 
 // src/exulu/context-guard.ts
@@ -16993,22 +17156,24 @@ var processFilePartsInMessages = async (messages, offloadCtx) => {
             return part;
           }
           console.log(`[EXULU] Processing part`, part);
-          const { mediaType, url, filename } = part;
-          console.log(`[EXULU] Media type: ${mediaType}`);
-          console.log(`[EXULU] URL: ${url}`);
-          console.log(`[EXULU] Filename: ${filename}`);
-          const imageTypes = [".png", ".jpeg", ".jpg", ".gif", ".webp"];
-          const imageType = imageTypes.find(
-            (type) => filename.toLowerCase().includes(type.toLowerCase())
-          );
-          if (imageType) {
-            console.log(`[EXULU] Converting file part to image part: ${filename} `);
+          const uploads = offloadCtx.exuluConfig?.fileUploads;
+          const url = uploads?.s3Bucket ? await resolveFreshFileUrl(part.url, {
+            endpoint: uploads.s3endpoint,
+            bucket: uploads.s3Bucket,
+            sign: (bucket, key) => getPresignedUrl(bucket, key, offloadCtx.exuluConfig)
+          }) : part.url;
+          const classified = classifyFilePart(part);
+          console.log(`[EXULU] File part classified as ${classified.kind}: ${classified.filename}`);
+          if (classified.kind === "image") {
             return {
               type: "file",
-              mediaType: `image/${imageType.replace(".", "")}`,
-              url
+              mediaType: classified.mediaType,
+              url,
+              // Keep the name so later turns/steps still know the file.
+              ...classified.filename ? { filename: classified.filename } : {}
             };
           }
+          const filename = classified.filename;
           console.log(`[EXULU] Converting file part to text using officeparser: ${filename}`);
           try {
             const response = await fetch(url);
@@ -17058,7 +17223,7 @@ var saveChat = async ({
   model
 }) => {
   const { db: db2 } = await postgresClient();
-  for (const message of messages) {
+  for (const message of dropEmptyMessages(messages)) {
     const mutation = db2.from("agent_messages").insert({
       session,
       user,
@@ -17145,7 +17310,7 @@ var generateSync = async ({
     );
     messages = await (0, import_ai6.validateUIMessages)({
       // append the new message to the previous messages:
-      messages: [...previousMessagesContent, ...messages]
+      messages: dropEmptyMessages([...previousMessagesContent, ...messages])
     });
     const contextBudget = deriveContextBudget(contextWindow);
     const occupancy = contextOccupancy(messages);
@@ -17487,7 +17652,7 @@ var generateStream = async ({
   const model = languageModel;
   messages = await (0, import_ai6.validateUIMessages)({
     // append the new message to the previous messages:
-    messages: [...previousMessagesContent, message]
+    messages: dropEmptyMessages([...previousMessagesContent, message])
   });
   const query = message.parts?.[0]?.type === "text" ? message.parts[0].text : void 0;
   if (session && query) {
@@ -19291,7 +19456,8 @@ var createWorkers = async (queues2, config, contexts, evals, tools, tracer) => {
           JOB_STATUS_ENUM.completed
         ]).update({
           state: JOB_STATUS_ENUM.failed,
-          error
+          // A raw Error serialises to "{}" in jsonb — keep message/stack/cause.
+          error: serializeError(error)
         });
         void maybePruneJobResults(db2);
         return;
@@ -23821,7 +23987,7 @@ var import_fs3 = __toESM(require("fs"), 1);
 var import_node_crypto15 = require("crypto");
 var import_api2 = require("@opentelemetry/api");
 var import_jszip3 = __toESM(require("jszip"), 1);
-var import_ai11 = require("ai");
+var import_ai12 = require("ai");
 var import_cookie_parser = __toESM(require("cookie-parser"), 1);
 init_statistics2();
 
@@ -23980,7 +24146,9 @@ var compactSession = async ({
 }) => {
   const budget = deriveContextBudget(contextWindow);
   const rows = await getAgentMessages({ session: sessionID, user: user.id });
-  const all = await (0, import_ai10.validateUIMessages)({ messages: rows.map((r) => JSON.parse(r.content)) });
+  const all = await (0, import_ai10.validateUIMessages)({
+    messages: dropEmptyMessages(rows.map((r) => JSON.parse(r.content)))
+  });
   const history = sliceHistoryAtCheckpoint(all);
   const { head, tail } = splitTail(history, budget.compactionTailTokens);
   if (head.length === 0) {
@@ -24034,6 +24202,25 @@ ${summary}` }],
   return { checkpoint, occupancyEstimate, originalTokens, summaryTokens };
 };
 
+// src/exulu/request-error.ts
+init_cjs_shims();
+var import_ai11 = require("ai");
+init_context_budget();
+function describeRequestError(err) {
+  if (err instanceof ContextCompactionRequiredError) {
+    return { status: 413, body: err.message };
+  }
+  if (import_ai11.TypeValidationError.isInstance(err)) {
+    const cause = err.cause;
+    const detail = cause instanceof Error ? cause.message : typeof cause === "string" ? cause : "";
+    return {
+      status: 422,
+      body: `The stored conversation could not be loaded: ${detail.slice(0, 400) || "type validation failed"}`
+    };
+  }
+  return { status: 500, body: err instanceof Error ? err.message : String(err) };
+}
+
 // src/exulu/transcribe.ts
 init_cjs_shims();
 init_env();
@@ -25639,11 +25826,13 @@ ${customInstructions}` : agent.instructions;
           });
         } catch (err) {
           if (headers.session) clearStreamActive(headers.session);
-          if (err instanceof ContextCompactionRequiredError) {
-            res.status(413).send(err.message);
-            return;
-          }
-          throw err;
+          console.error(
+            "[EXULU] chat request failed before streaming.",
+            err instanceof Error ? err.message.slice(0, 500) : err
+          );
+          const { status, body } = describeRequestError(err);
+          res.status(status).send(body);
+          return;
         }
         result.stream.consumeStream();
         result.stream.pipeUIMessageStreamToResponse(res, {
@@ -25678,7 +25867,7 @@ ${customInstructions}` : agent.instructions;
             else message2 = JSON.stringify(error);
             return mapStreamErrorMessage(message2);
           },
-          generateMessageId: (0, import_ai11.createIdGenerator)({
+          generateMessageId: (0, import_ai12.createIdGenerator)({
             prefix: "msg_",
             size: 16
           }),
@@ -25782,11 +25971,13 @@ ${customInstructions}` : agent.instructions;
             }
           });
         } catch (err) {
-          if (err instanceof ContextCompactionRequiredError) {
-            res.status(413).send(err.message);
-            return;
-          }
-          throw err;
+          console.error(
+            "[EXULU] run request failed.",
+            err instanceof Error ? err.message.slice(0, 500) : err
+          );
+          const { status, body } = describeRequestError(err);
+          res.status(status).send(body);
+          return;
         }
         res.status(200).json(response);
         return;
@@ -28110,6 +28301,7 @@ ${style.markdown}` : params.prompt;
     ".jpeg": "image/jpeg",
     ".gif": "image/gif",
     ".webp": "image/webp",
+    ".bmp": "image/bmp",
     ".svg": "image/svg+xml",
     ".pdf": "application/pdf",
     ".docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",
@@ -29136,7 +29328,7 @@ init_cjs_shims();
 
 // src/exulu/evals.ts
 init_cjs_shims();
-var import_ai12 = require("ai");
+var import_ai13 = require("ai");
 var ExuluEval = class {
   id;
   name;
@@ -29176,7 +29368,7 @@ var ExuluEval = class {
 init_resolve_model();
 init_singleton();
 var import_zod20 = require("zod");
-var import_ai13 = require("ai");
+var import_ai14 = require("ai");
 var llmAsJudgeEval = () => {
   if (process.env.REDIS_HOST?.length && process.env.REDIS_PORT?.length) {
     return new ExuluEval({
@@ -29219,13 +29411,13 @@ var llmAsJudgeEval = () => {
           rbacBypass: true
         });
         console.log("[EXULU] prompt", prompt);
-        const { output } = await (0, import_ai13.generateText)({
+        const { output } = await (0, import_ai14.generateText)({
           temperature: 0,
           model: resolved.languageModel,
           system: "",
           prompt,
           maxRetries: 2,
-          output: import_ai13.Output.object({
+          output: import_ai14.Output.object({
             schema: import_zod20.z.object({
               score: import_zod20.z.number().min(0).max(100).describe("The score between 0 and 100.")
             })
@@ -32839,7 +33031,7 @@ var MarkdownChunker = class {
 init_cjs_shims();
 var fs4 = __toESM(require("fs"), 1);
 var path3 = __toESM(require("path"), 1);
-var import_ai14 = require("ai");
+var import_ai15 = require("ai");
 var import_zod27 = require("zod");
 var import_p_limit = __toESM(require("p-limit"), 1);
 var import_crypto3 = require("crypto");
@@ -33295,9 +33487,9 @@ If the page contains a flow-chart, schematic, technical drawing or control board
 
 ### 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_ai14.generateText)({
+  const result = await (0, import_ai15.generateText)({
     model,
-    output: import_ai14.Output.object({
+    output: import_ai15.Output.object({
       schema: import_zod27.z.object({
         needs_correction: import_zod27.z.boolean(),
         corrected_text: import_zod27.z.string().nullable(),

package/dist/index.js

@@ -1347,17 +1347,31 @@ function detectQueryLanguage(query, minLength = 10) {
 }
 function stemWord(word, languageCode) {
   const stemmer = STEMMER_MAP[languageCode] || natural.PorterStemmer;
-  const cleaned = word.replace(/[^\p{L}\p{N}]/gu, "").toLowerCase();
-  if (!cleaned) {
+  const core = word.replace(/^[^\p{L}\p{N}]+/u, "").replace(/[^\p{L}\p{N}]+$/u, "").toLowerCase();
+  if (!core) {
     return word;
   }
+  if (!/^\p{L}+$/u.test(core)) {
+    return core;
+  }
   try {
-    return stemmer.stem(cleaned);
+    return stemmer.stem(core);
   } catch (error) {
     console.warn(`[EXULU] Error stemming word "${word}":`, error);
-    return cleaned;
+    return core;
   }
 }
+function buildFullTextOrQuery(original, processed) {
+  const seen = /* @__PURE__ */ new Set();
+  const tokens = [];
+  for (const raw of `${original} ${processed}`.split(/\s+/)) {
+    const token = raw.replace(/["'`]/g, "").replace(/^[^\p{L}\p{N}]+/u, "").replace(/[^\p{L}\p{N}]+$/u, "").toLowerCase();
+    if (!token || token === "or" || seen.has(token)) continue;
+    seen.add(token);
+    tokens.push(token);
+  }
+  return tokens.join(" or ");
+}
 function preprocessQuery(query, options = {}) {
   const {
     enableStemming = true,
@@ -1392,6 +1406,18 @@ function preprocessQuery(query, options = {}) {
     stemmed: true
   };
 }
+function resolveSearchQueryTexts(query) {
+  const { processed } = preprocessQuery(query, {
+    enableStemming: true,
+    detectLanguage: true
+  });
+  const ftsText = processed || query;
+  return {
+    embedText: query,
+    ftsText,
+    hybridOrQuery: buildFullTextOrQuery(query, ftsText)
+  };
+}
 
 // src/graphql/resolvers/field-allow-list.ts
 var ALWAYS_ALLOWED = /* @__PURE__ */ new Set(["id", "createdAt", "updatedAt"]);
@@ -3828,17 +3854,15 @@ var vectorSearch = async ({
   const _tBody = Date.now();
   let _preMs = 0, _statMs = 0, _resolveMs = 0, _embedMs = 0;
   let _embedSource = "none";
+  let hybridOrQuery = "";
   if (query) {
     const _tp = Date.now();
-    const { processed: stemmedQuery } = preprocessQuery(query, {
-      enableStemming: true,
-      detectLanguage: true
-    });
+    const texts = resolveSearchQueryTexts(query);
     _preMs = Date.now() - _tp;
-    console.log("[EXULU] Stemmed query:", stemmedQuery);
-    if (stemmedQuery) {
-      query = stemmedQuery;
-    }
+    console.log("[EXULU] Search query texts:", texts);
+    const embedText = texts.embedText;
+    hybridOrQuery = texts.hybridOrQuery;
+    query = texts.ftsText;
     if (queryEmbedding && queryEmbedding.length) {
       vector = queryEmbedding;
       _embedSource = "reused";
@@ -3864,7 +3888,7 @@ var vectorSearch = async ({
       });
       _resolveMs = Date.now() - _tr;
       const _te = Date.now();
-      const [queryVector] = await resolved.embed([query], { inputType: "query" });
+      const [queryVector] = await resolved.embed([embedText], { inputType: "query" });
       _embedMs = Date.now() - _te;
       if (!queryVector?.length) {
         throw new Error("No vector generated for query.");
@@ -3936,10 +3960,10 @@ var vectorSearch = async ({
       const fullTextWeight = 2;
       const semanticWeight = 1;
       const rrfK = 50;
-      const ftRankExpression = languages.map((lang) => `ts_rank(chunks.fts, plainto_tsquery('${lang}', ?))`).join(", ");
-      const ftRankParams = languages.map(() => query);
-      const ftMatchExpression = languages.map((lang) => `chunks.fts @@ plainto_tsquery('${lang}', ?)`).join(" OR ");
-      const ftMatchParams = languages.map(() => query);
+      const ftRankExpression = languages.map((lang) => `ts_rank(chunks.fts, websearch_to_tsquery('${lang}', ?))`).join(", ");
+      const ftRankParams = languages.map(() => hybridOrQuery);
+      const ftMatchExpression = languages.map((lang) => `chunks.fts @@ websearch_to_tsquery('${lang}', ?)`).join(" OR ");
+      const ftMatchParams = languages.map(() => hybridOrQuery);
       let fullTextQuery = db(chunksTable + " as chunks").select([
         "chunks.id",
         "chunks.source",
@@ -3979,8 +4003,8 @@ var vectorSearch = async ({
         db.raw('items."updatedAt" as item_updated_at'),
         db.raw('items."createdAt" as item_created_at'),
         db.raw(
-          `GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, plainto_tsquery('${lang}', ?))`).join(", ")}) AS fts_rank`,
-          languages.map(() => query)
+          `GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, websearch_to_tsquery('${lang}', ?))`).join(", ")}) AS fts_rank`,
+          languages.map(() => hybridOrQuery)
         ),
         db.raw(`(1 - (chunks.embedding <=> ${vectorExpr})) AS cosine_distance`),
         db.raw(
@@ -4005,8 +4029,8 @@ var vectorSearch = async ({
                   `,
         [rrfK, fullTextWeight, rrfK, semanticWeight, cutoffs?.hybrid || 0]
       ).whereRaw(
-        `(chunks.fts IS NULL OR GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, plainto_tsquery('${lang}', ?))`).join(", ")}) > ?)`,
-        [...languages.map(() => query), cutoffs?.tsvector || 0]
+        `(chunks.fts IS NULL OR GREATEST(${languages.map((lang) => `ts_rank(chunks.fts, websearch_to_tsquery('${lang}', ?))`).join(", ")}) > ?)`,
+        [...languages.map(() => hybridOrQuery), cutoffs?.tsvector || 0]
       ).whereRaw(`(chunks.embedding IS NULL OR (1 - (chunks.embedding <=> ${vectorExpr})) >= ?)`, [
         cutoffs?.cosineDistance || 0
       ]).orderByRaw("hybrid_score DESC").limit(Math.min(matchCount, 250));
@@ -7687,9 +7711,141 @@ var bullmq = {
   }
 };
 
+// src/utils/serialize-error.ts
+var MAX_CAUSE_DEPTH = 3;
+function serializeError(err, depth = 0) {
+  if (err instanceof Error) {
+    const out = { name: err.name, message: err.message };
+    if (err.stack) out.stack = err.stack;
+    for (const [key, value] of Object.entries(err)) {
+      if (key !== "cause") out[key] = value;
+    }
+    const cause = err.cause;
+    if (cause !== void 0 && depth < MAX_CAUSE_DEPTH) {
+      out.cause = serializeError(cause, depth + 1);
+    }
+    return out;
+  }
+  if (typeof err === "string") {
+    return { message: err };
+  }
+  if (err !== null && typeof err === "object") {
+    let message;
+    try {
+      message = JSON.stringify(err);
+    } catch {
+      message = "[unserializable error object]";
+    }
+    return { ...err, message };
+  }
+  return { message: String(err) };
+}
+
 // src/exulu/agent-as-tool.ts
 import { z as z2 } from "zod";
 
+// src/exulu/file-part.ts
+var IMAGE_EXTENSION_MIME = {
+  ".png": "image/png",
+  ".jpeg": "image/jpeg",
+  ".jpg": "image/jpeg",
+  ".gif": "image/gif",
+  ".webp": "image/webp",
+  // Windows bitmaps arrive as email logos / signature images and camera exports.
+  // Left out, they were routed to OfficeParser and replaced by an error text.
+  ".bmp": "image/bmp",
+  ".dib": "image/bmp"
+};
+var MEDIA_TYPE_ALIASES = {
+  "image/jpg": "image/jpeg",
+  "image/x-ms-bmp": "image/bmp",
+  "image/x-bmp": "image/bmp",
+  "image/x-windows-bmp": "image/bmp"
+};
+function fileNameFromUrl(url) {
+  if (!url) return void 0;
+  let path2;
+  try {
+    path2 = new URL(url).pathname;
+  } catch {
+    path2 = url.split("?")[0] ?? "";
+  }
+  const last = path2.split("/").pop() ?? "";
+  try {
+    return decodeURIComponent(last) || void 0;
+  } catch {
+    return last || void 0;
+  }
+}
+function classifyFilePart(part) {
+  const filename = part.filename ?? fileNameFromUrl(part.url);
+  const lowerName = (filename ?? "").toLowerCase();
+  const extension = Object.keys(IMAGE_EXTENSION_MIME).find((ext) => lowerName.endsWith(ext));
+  const declared = typeof part.mediaType === "string" ? part.mediaType.trim().toLowerCase() : "";
+  const declaredIsImage = declared.startsWith("image/");
+  if (declaredIsImage || extension) {
+    const mediaType = declaredIsImage ? MEDIA_TYPE_ALIASES[declared] ?? declared : IMAGE_EXTENSION_MIME[extension];
+    return { kind: "image", mediaType, filename };
+  }
+  return { kind: "document", filename: filename ?? "attachment" };
+}
+
+// src/exulu/stored-messages.ts
+function dropEmptyMessages(messages) {
+  const kept = messages.filter((m) => Array.isArray(m.parts) && m.parts.length > 0);
+  return kept.length === messages.length ? messages : kept;
+}
+
+// src/exulu/stored-file-url.ts
+var stripTrailingSlash = (s) => s.replace(/\/+$/, "");
+function decodeKey(rawPath) {
+  const key = rawPath.replace(/^\/+/, "");
+  try {
+    return decodeURIComponent(key);
+  } catch {
+    return key;
+  }
+}
+function parseStoredFileUrl(url, store) {
+  if (!url || !store.bucket) return void 0;
+  let parsed;
+  try {
+    parsed = new URL(url);
+  } catch {
+    return void 0;
+  }
+  if (parsed.protocol !== "http:" && parsed.protocol !== "https:") return void 0;
+  if (store.endpoint) {
+    let endpoint;
+    try {
+      endpoint = new URL(store.endpoint);
+    } catch {
+      return void 0;
+    }
+    if (parsed.host !== endpoint.host) return void 0;
+    const base = stripTrailingSlash(endpoint.pathname);
+    const prefix = `${base}/${store.bucket}/`;
+    if (!parsed.pathname.startsWith(prefix)) return void 0;
+    const key2 = decodeKey(parsed.pathname.slice(prefix.length));
+    return key2 ? { bucket: store.bucket, key: key2 } : void 0;
+  }
+  if (!parsed.host.startsWith(`${store.bucket}.s3`) || !parsed.host.endsWith(".amazonaws.com")) {
+    return void 0;
+  }
+  const key = decodeKey(parsed.pathname);
+  return key ? { bucket: store.bucket, key } : void 0;
+}
+async function resolveFreshFileUrl(url, opts) {
+  const location = parseStoredFileUrl(url, opts);
+  if (!location) return url;
+  try {
+    return await opts.sign(location.bucket, location.key);
+  } catch (err) {
+    console.warn(`[EXULU] could not re-sign stored file URL for key "${location.key}":`, err);
+    return url;
+  }
+}
+
 // src/exulu/generate-stream.ts
 import {
   convertToModelMessages,
@@ -7993,22 +8149,24 @@ var processFilePartsInMessages = async (messages, offloadCtx) => {
             return part;
           }
           console.log(`[EXULU] Processing part`, part);
-          const { mediaType, url, filename } = part;
-          console.log(`[EXULU] Media type: ${mediaType}`);
-          console.log(`[EXULU] URL: ${url}`);
-          console.log(`[EXULU] Filename: ${filename}`);
-          const imageTypes = [".png", ".jpeg", ".jpg", ".gif", ".webp"];
-          const imageType = imageTypes.find(
-            (type) => filename.toLowerCase().includes(type.toLowerCase())
-          );
-          if (imageType) {
-            console.log(`[EXULU] Converting file part to image part: ${filename} `);
+          const uploads = offloadCtx.exuluConfig?.fileUploads;
+          const url = uploads?.s3Bucket ? await resolveFreshFileUrl(part.url, {
+            endpoint: uploads.s3endpoint,
+            bucket: uploads.s3Bucket,
+            sign: (bucket, key) => getPresignedUrl(bucket, key, offloadCtx.exuluConfig)
+          }) : part.url;
+          const classified = classifyFilePart(part);
+          console.log(`[EXULU] File part classified as ${classified.kind}: ${classified.filename}`);
+          if (classified.kind === "image") {
             return {
               type: "file",
-              mediaType: `image/${imageType.replace(".", "")}`,
-              url
+              mediaType: classified.mediaType,
+              url,
+              // Keep the name so later turns/steps still know the file.
+              ...classified.filename ? { filename: classified.filename } : {}
             };
           }
+          const filename = classified.filename;
           console.log(`[EXULU] Converting file part to text using officeparser: ${filename}`);
           try {
             const response = await fetch(url);
@@ -8058,7 +8216,7 @@ var saveChat = async ({
   model
 }) => {
   const { db } = await postgresClient();
-  for (const message of messages) {
+  for (const message of dropEmptyMessages(messages)) {
     const mutation = db.from("agent_messages").insert({
       session,
       user,
@@ -8145,7 +8303,7 @@ var generateSync = async ({
     );
     messages = await validateUIMessages({
       // append the new message to the previous messages:
-      messages: [...previousMessagesContent, ...messages]
+      messages: dropEmptyMessages([...previousMessagesContent, ...messages])
     });
     const contextBudget = deriveContextBudget(contextWindow);
     const occupancy = contextOccupancy(messages);
@@ -8487,7 +8645,7 @@ var generateStream = async ({
   const model = languageModel;
   messages = await validateUIMessages({
     // append the new message to the previous messages:
-    messages: [...previousMessagesContent, message]
+    messages: dropEmptyMessages([...previousMessagesContent, message])
   });
   const query = message.parts?.[0]?.type === "text" ? message.parts[0].text : void 0;
   if (session && query) {
@@ -10270,7 +10428,8 @@ var createWorkers = async (queues2, config, contexts, evals, tools, tracer) => {
           JOB_STATUS_ENUM.completed
         ]).update({
           state: JOB_STATUS_ENUM.failed,
-          error
+          // A raw Error serialises to "{}" in jsonb — keep message/stack/cause.
+          error: serializeError(error)
         });
         void maybePruneJobResults(db);
         return;
@@ -14914,7 +15073,9 @@ var compactSession = async ({
 }) => {
   const budget = deriveContextBudget(contextWindow);
   const rows = await getAgentMessages({ session: sessionID, user: user.id });
-  const all = await validateUIMessages2({ messages: rows.map((r) => JSON.parse(r.content)) });
+  const all = await validateUIMessages2({
+    messages: dropEmptyMessages(rows.map((r) => JSON.parse(r.content)))
+  });
   const history = sliceHistoryAtCheckpoint(all);
   const { head, tail } = splitTail(history, budget.compactionTailTokens);
   if (head.length === 0) {
@@ -14968,6 +15129,23 @@ ${summary}` }],
   return { checkpoint, occupancyEstimate, originalTokens, summaryTokens };
 };
 
+// src/exulu/request-error.ts
+import { TypeValidationError } from "ai";
+function describeRequestError(err) {
+  if (err instanceof ContextCompactionRequiredError) {
+    return { status: 413, body: err.message };
+  }
+  if (TypeValidationError.isInstance(err)) {
+    const cause = err.cause;
+    const detail = cause instanceof Error ? cause.message : typeof cause === "string" ? cause : "";
+    return {
+      status: 422,
+      body: `The stored conversation could not be loaded: ${detail.slice(0, 400) || "type validation failed"}`
+    };
+  }
+  return { status: 500, body: err instanceof Error ? err.message : String(err) };
+}
+
 // src/exulu/transcribe.ts
 var TRANSCRIBE_SYSTEM_PROMPT = "You are a speech-to-text transcription engine. Detect the language actually spoken and transcribe it word-for-word in that same language. Never translate. Output only the transcript text \u2014 no quotes, labels, or commentary. If there is no intelligible speech, output nothing.";
 function isGeminiChatTranscriptionModel(entry) {
@@ -16524,11 +16702,13 @@ ${customInstructions}` : agent.instructions;
           });
         } catch (err) {
           if (headers.session) clearStreamActive(headers.session);
-          if (err instanceof ContextCompactionRequiredError) {
-            res.status(413).send(err.message);
-            return;
-          }
-          throw err;
+          console.error(
+            "[EXULU] chat request failed before streaming.",
+            err instanceof Error ? err.message.slice(0, 500) : err
+          );
+          const { status, body } = describeRequestError(err);
+          res.status(status).send(body);
+          return;
         }
         result.stream.consumeStream();
         result.stream.pipeUIMessageStreamToResponse(res, {
@@ -16667,11 +16847,13 @@ ${customInstructions}` : agent.instructions;
             }
           });
         } catch (err) {
-          if (err instanceof ContextCompactionRequiredError) {
-            res.status(413).send(err.message);
-            return;
-          }
-          throw err;
+          console.error(
+            "[EXULU] run request failed.",
+            err instanceof Error ? err.message.slice(0, 500) : err
+          );
+          const { status, body } = describeRequestError(err);
+          res.status(status).send(body);
+          return;
         }
         res.status(200).json(response);
         return;
@@ -18995,6 +19177,7 @@ ${style.markdown}` : params.prompt;
     ".jpeg": "image/jpeg",
     ".gif": "image/gif",
     ".webp": "image/webp",
+    ".bmp": "image/bmp",
     ".svg": "image/svg+xml",
     ".pdf": "application/pdf",
     ".docx": "application/vnd.openxmlformats-officedocument.wordprocessingml.document",

package/ee/workers.ts

@@ -3,6 +3,7 @@ import { redisServer } from "@EE/queues/server.ts";
 import { guardRedisStartup, logRedisErrors } from "@EE/queues/redis-startup.ts";
 import { Job, Worker, type JobState } from "bullmq";
 import { bullmq } from "@SRC/validators/bullmq.ts";
+import { serializeError } from "@SRC/utils/serialize-error.ts";
 import { getEnabledTools } from "@SRC/utils/enabled-tools.ts";
 import { ExuluStorage } from "@SRC/exulu/storage.ts";
 import type { ExuluAgent } from "@EXULU_TYPES/models/agent.ts";
@@ -1212,7 +1213,8 @@ export const createWorkers = async (
           ])
           .update({
             state: JOB_STATUS_ENUM.failed,
-            error,
+            // A raw Error serialises to "{}" in jsonb — keep message/stack/cause.
+            error: serializeError(error),
           });
 
         // Cap the table as rows become terminal (every Nth, idempotent).

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@exulu/backend",
   "author": "Qventu Bv.",
-  "version": "3.7.1",
+  "version": "3.7.3",
   "main": "./dist/index.js",
   "private": false,
   "publishConfig": {