@exulu/backend 1.58.0 → 1.59.0

package/ee/python/.litellm/config.yaml.example

@@ -0,0 +1,64 @@
+# LiteLLM Proxy configuration for Exulu.
+#
+# This file is read by the LiteLLM process Exulu spawns when EXULU_USE_LITELLM=true.
+# Documentation: https://docs.litellm.ai/docs/proxy/configs
+#
+# To use:
+#   1. Copy this file to config.yaml in the same directory.
+#   2. Edit model_list to add the providers you want to expose.
+#   3. Set the referenced env vars (LITELLM_MASTER_KEY, GOOGLE_VERTEX_PROJECT, etc.).
+#   4. Set EXULU_USE_LITELLM=true and (re)start Exulu.
+
+general_settings:
+  master_key: os.environ/LITELLM_MASTER_KEY
+
+  # Database mode (optional). Set `database_url` only if you want LiteLLM's
+  # user/key/budget/spend tracking tables. When set, LiteLLM creates ~60
+  # tables in the target database (all prefixed `LiteLLM_`).
+  #
+  # ⚠ ALWAYS use a DEDICATED Postgres database for LiteLLM — never share it
+  #   with Exulu's main database. LiteLLM's schema sync drops any tables in
+  #   the public schema that aren't part of its schema; sharing a database
+  #   with Exulu has destroyed application data in the past.
+  #
+  # Exulu's ExuluDatabase.init runs `prisma db push` against this URL
+  # automatically at boot, but only after verifying:
+  #   1. database_url is NOT the same as Exulu's own Postgres (warns + skips)
+  #   2. the target database contains no non-LiteLLM_* tables (warns + skips)
+  #   3. the push itself runs without --accept-data-loss (Prisma refuses any
+  #      schema change that would drop data)
+  # Re-running on an already-set-up database is a no-op.
+  #
+  # database_url: "postgresql://user:pass@host:5432/litellm_dedicated"
+
+model_list:
+  # Each entry's `model_name` is what Exulu agents will reference in their
+  # `model` field after toggling LiteLLM on. Pick stable, human-friendly names.
+
+  - model_name: vertex-flash
+    litellm_params:
+      model: vertex_ai/gemini-2.5-flash
+      vertex_project: os.environ/GOOGLE_VERTEX_PROJECT
+      vertex_location: europe-west1
+      vertex_credentials: os.environ/GOOGLE_VERTEX_CREDENTIALS_JSON
+
+  - model_name: vertex-pro
+    litellm_params:
+      model: vertex_ai/gemini-2.5-pro
+      vertex_project: os.environ/GOOGLE_VERTEX_PROJECT
+      vertex_location: europe-west1
+      vertex_credentials: os.environ/GOOGLE_VERTEX_CREDENTIALS_JSON
+
+  - model_name: claude-haiku
+    litellm_params:
+      model: anthropic/claude-haiku-4-5
+      api_key: os.environ/ANTHROPIC_API_KEY
+
+  - model_name: gpt-5
+    litellm_params:
+      model: openai/gpt-5
+      api_key: os.environ/OPENAI_API_KEY
+
+# Per-model rate limits and budgets are configured here in LiteLLM-native form.
+# See https://docs.litellm.ai/docs/proxy/users for budget configuration and
+# https://docs.litellm.ai/docs/proxy/rate_limit_tiers for rate limiting.

package/ee/python/requirements.txt

@@ -3,3 +3,18 @@ transformers
 pyinstaller
 docling-hierarchical-pdf
 defusedxml
+# LiteLLM proxy — only used when EXULU_USE_LITELLM=true. Always installed so
+# the dep is ready when the env var is flipped. Pinned to a tested version;
+# upgrade deliberately.
+litellm[proxy]==1.85.1
+# Prisma Python client — required by LiteLLM proxy whenever a `database_url`
+# is set in config.litellm.yaml (used for user/budget tracking, key
+# management, etc.). NOT included in litellm[proxy] in 1.85.1, hence the
+# separate pin. setup.sh runs `prisma generate` against LiteLLM's bundled
+# schema after pip install so the Python client module is materialized.
+prisma==0.15.0
+# Vertex AI SDK — required by LiteLLM when routing to `vertex_ai/*` models
+# (the `vertexai` Python module lives inside google-cloud-aiplatform).
+# NOT included in litellm[proxy]; without it, requests to Vertex models
+# fail with "No module named 'vertexai'".
+google-cloud-aiplatform>=1.38

package/ee/python/setup.sh

@@ -203,6 +203,19 @@ pip install -r "$REQUIREMENTS_FILE"
 
 print_success "All dependencies installed successfully"
 
+# Step 6.5: Generate Prisma client for LiteLLM database mode.
+# LiteLLM's PrismaClient does `from prisma import Prisma`, which only works
+# after `prisma generate` has materialized the Python client against
+# LiteLLM's bundled schema. Skip silently if LiteLLM isn't installed or its
+# schema isn't where we expect — database mode is opt-in via config.litellm.yaml.
+LITELLM_PROXY_DIR=$(find "$VENV_DIR/lib" -path "*/litellm/proxy" -type d 2>/dev/null | head -1)
+if [ -n "$LITELLM_PROXY_DIR" ] && [ -f "$LITELLM_PROXY_DIR/schema.prisma" ]; then
+    print_info "Generating Prisma client for LiteLLM..."
+    (cd "$LITELLM_PROXY_DIR" && PATH="$VENV_DIR/bin:$PATH" "$VENV_DIR/bin/prisma" generate > /dev/null 2>&1) \
+        && print_success "Prisma client generated for LiteLLM" \
+        || print_warning "Prisma generate failed; LiteLLM database mode (database_url in config.litellm.yaml) may not work until you run 'cd $LITELLM_PROXY_DIR && PATH=$VENV_DIR/bin:\$PATH $VENV_DIR/bin/prisma generate'"
+fi
+
 # Step 7: Validate installation
 echo ""
 echo "Step 7: Validating installation..."

package/ee/workers.ts

@@ -10,6 +10,7 @@ import { getTableName, type ExuluContext } from "@SRC/exulu/context.ts";
 import type { ExuluReranker } from "@SRC/exulu/reranker.ts";
 import type { ExuluEval } from "@SRC/exulu/evals.ts";
 import type { ExuluTool } from "@SRC/exulu/tool.ts";
+import { resolveModel } from "@SRC/exulu/resolve-model.ts";
 import { postgresClient } from "@SRC/postgres/client";
 import type { BullMqJobData } from "@EE/queues/decorator.ts";
 import { type Tracer } from "@opentelemetry/api";
@@ -1379,37 +1380,21 @@ export const processUiMessagesFlow = async ({
     enabledTools?.map((x) => x.name + " (" + x.id + ")"),
   );
 
-  // Get the variable name from user's anthropic_token field
-  const variableName = agent.providerapikey;
-
-  // Look up the variable from the variables table
-  const { db } = await postgresClient();
-
-  let providerapikey: string | undefined;
-
-  if (variableName) {
-    const variable = await db.from("variables").where({ name: variableName }).first();
-    if (!variable) {
-      throw new Error(
-        `Provider API key variable not found for agent ${agent.name} (${agent.id}).`,
-      );
-    }
-
-    // Get the API key from the variable (decrypt if encrypted)
-    providerapikey = variable.value;
-
-    if (!variable.encrypted) {
-      throw new Error(
-        `Provider API key variable not encrypted for agent ${agent.name} (${agent.id}), for security reasons you are only allowed to use encrypted variables for provider API keys.`,
-      );
-    }
-
-    if (variable.encrypted) {
-      const bytes = CryptoJS.AES.decrypt(variable.value, process.env.NEXTAUTH_SECRET);
-      providerapikey = bytes.toString(CryptoJS.enc.Utf8);
-    }
+  if (!agent.model) {
+    throw new Error(
+      `Agent ${agent.name} (${agent.id}) has no model configured.`,
+    );
   }
 
+  const resolved = await resolveModel({
+    modelId: agent.model,
+    user,
+    providers,
+    agent: { id: agent.id },
+  });
+  const providerapikey = resolved.apiKey;
+  const resolvedLanguageModel = resolved.languageModel;
+
   // Remove placeholder agent response before sending
   const messagesWithoutPlaceholder = inputMessages.filter(
     (message) => (message.metadata as any)?.type !== "placeholder",
@@ -1512,6 +1497,7 @@ export const processUiMessagesFlow = async ({
           message: currentMessage,
           currentTools: enabledTools,
           allExuluTools: tools,
+          languageModel: resolvedLanguageModel,
           providerapikey,
           toolConfigs: agent.tools,
           exuluConfig: config,

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@exulu/backend",
   "author": "Qventu Bv.",
-  "version": "1.58.0",
+  "version": "1.59.0",
   "main": "./dist/index.js",
   "private": false,
   "publishConfig": {
@@ -63,6 +63,7 @@
     "@types/express": "^5.0.1",
     "@types/graphql-type-json": "^0.3.5",
     "@types/jest": "^29.5.14",
+    "@types/multer": "^2.1.0",
     "@types/node": "^22.14.0",
     "@types/pg": "^8.15.1",
     "@types/supertest": "^6.0.2",
@@ -139,6 +140,7 @@
     "knex": "^3.1.0",
     "link": "^2.1.1",
     "mammoth": "^1.11.0",
+    "multer": "^2.1.1",
     "natural": "^8.1.0",
     "nodemailer": "^8.0.7",
     "officeparser": "^5.2.2",