LCORE-2860: fix RHOAI Prow e2e pipeline failures

- Remove --image flag from oc new-build that overrode the containerfile's
  FROM (ubi9/python-312) with ubi-minimal, which lacks dnf
- Remove duplicate vector_stores registration from run.yaml (enrichment
  handles it)
- Add RHOAI-specific lightspeed-stack config with vLLM defaults
- Parametrize model name via vllm-model-secret so both pods read from
  a single source
- Override inference defaults (default_model/default_provider) in Prow
  config swaps using E2E env vars
- Create /tmp/data before starting lightspeed-stack pod for SQLite

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: are-ces

tests/e2e-prow/rhoai/configs/run.yaml

@@ -139,17 +139,7 @@ registered_resources:
     provider_id: vllm
     model_type: llm
     provider_model_id: null
-  - model_id: all-mpnet-base-v2
-    model_type: embedding
-    provider_id: sentence-transformers
-    provider_model_id: all-mpnet-base-v2
-    metadata:
-      embedding_dimension: 768
-  vector_stores: 
-  - embedding_dimension: 768
-    embedding_model: sentence-transformers/all-mpnet-base-v2
-    provider_id: faiss
-    vector_store_id: ${env.FAISS_VECTOR_STORE_ID}
+  vector_stores: []
   shields:
   - shield_id: llama-guard
     provider_id: llama-guard

tests/e2e-prow/rhoai/manifests/lightspeed/lightspeed-stack.yaml

@@ -37,6 +37,7 @@ spec:
               key: key
               optional: true
       image: ${LIGHTSPEED_STACK_IMAGE}
+      command: ["/bin/sh", "-c", "mkdir -p /tmp/data && exec python3.12 src/lightspeed_stack.py"]
       ports:
         - containerPort: 8080
       # TCP probes avoid HTTP/auth. LCS + Llama handshake and large images can take 60–120s before :8080 listens;

tests/e2e-prow/rhoai/manifests/lightspeed/llama-stack-prow.yaml

@@ -102,7 +102,10 @@ spec:
               name: vllm-api-key-secret
               key: key
         - name: INFERENCE_MODEL
-          value: "meta-llama/Llama-3.1-8B-Instruct"
+          valueFrom:
+            secretKeyRef:
+              name: vllm-model-secret
+              key: key
         - name: OPENAI_API_KEY
           valueFrom:
             secretKeyRef:

tests/e2e-prow/rhoai/pipeline.sh

@@ -64,6 +64,7 @@ create_secret() {
 create_secret hf-token-secret --from-literal=token="$HUGGING_FACE_HUB_TOKEN"
 create_secret vllm-api-key-secret --from-literal=key="$VLLM_API_KEY"
 create_secret openai-api-key-secret --from-literal=key=""
+create_secret vllm-model-secret --from-literal=key="$MODEL_NAME"
 
 # MCP token secrets for lightspeed-stack
 REPO_ROOT="$(cd "$PIPELINE_DIR/../../.." && pwd)"
@@ -225,7 +226,6 @@ export LLAMA_STACK_IMAGE
 oc new-build --name=llama-stack-e2e \
   --binary \
   --strategy=docker \
-  --image="registry.access.redhat.com/ubi9/ubi-minimal" \
   --to="llama-stack-e2e:latest" \
   -n "$NAMESPACE" 2>/dev/null || echo "BuildConfig llama-stack-e2e already exists"
 
@@ -256,7 +256,7 @@ create_secret api-url-secret --from-literal=key="$KSVC_URL"
 oc create configmap llama-stack-config -n "$NAMESPACE" \
   --from-file="$REPO_ROOT/tests/e2e-prow/rhoai/configs/run.yaml"
 oc create configmap lightspeed-stack-config -n "$NAMESPACE" \
-  --from-file="$REPO_ROOT/tests/e2e/configuration/server-mode/lightspeed-stack.yaml"
+  --from-file=lightspeed-stack.yaml="$REPO_ROOT/tests/e2e/configuration/server-mode/lightspeed-stack-rhoai.yaml"
 
 # Create RAG data ConfigMap from the e2e test RAG data
 echo "Creating RAG data ConfigMap..."

tests/e2e/configuration/server-mode/lightspeed-stack-rhoai.yaml

@@ -0,0 +1,35 @@
+name: Lightspeed Core Service (LCS)
+service:
+  host: 0.0.0.0
+  port: 8080
+  auth_enabled: false
+  workers: 1
+  color_log: true
+  access_log: true
+llama_stack:
+  # Server mode - connects to separate llama-stack service
+  use_as_library_client: false
+  url: http://${env.E2E_LLAMA_HOSTNAME}:8321
+  api_key: xyzzy
+user_data_collection:
+  feedback_enabled: true
+  feedback_storage: "/tmp/data/feedback"
+  transcripts_enabled: true
+  transcripts_storage: "/tmp/data/transcripts"
+authentication:
+  module: "noop"
+inference:
+  default_provider: vllm
+  default_model: ${env.VLLM_MODEL}
+byok_rag:
+  - rag_id: e2e-test-docs
+    rag_type: inline::faiss
+    embedding_model: sentence-transformers/all-mpnet-base-v2
+    embedding_dimension: 768
+    vector_db_id: ${env.FAISS_VECTOR_STORE_ID}
+    db_path: ${env.KV_RAG_PATH:=~/.llama/storage/rag/kv_store.db}
+    score_multiplier: 1.0
+
+rag:
+  tool:
+    - e2e-test-docs

tests/e2e/utils/prow_utils.py

@@ -5,6 +5,7 @@
 """
 
 import os
+import re
 import subprocess
 import tempfile
 from typing import Optional
@@ -234,6 +235,24 @@ def remove_configmap_backup(backup_key: str) -> None:
         print(f"ConfigMap backup {backup_key} removed from memory")
 
 
+def _apply_inference_overrides(content: str) -> str:
+    """Override inference defaults from E2E env vars if set.
+
+    When E2E_DEFAULT_MODEL_OVERRIDE or E2E_DEFAULT_PROVIDER_OVERRIDE are set,
+    patch the YAML content so every config swap uses the correct model/provider
+    for the current environment (e.g. vLLM instead of OpenAI in RHOAI Prow).
+    """
+    model = os.getenv("E2E_DEFAULT_MODEL_OVERRIDE")
+    provider = os.getenv("E2E_DEFAULT_PROVIDER_OVERRIDE")
+    if not model and not provider:
+        return content
+    if model:
+        content = re.sub(r"(default_model:\s*).*", rf"\g<1>{model}", content)
+    if provider:
+        content = re.sub(r"(default_provider:\s*).*", rf"\g<1>{provider}", content)
+    return content
+
+
 def _recreate_configmap(
     configmap_name: str,
     source_file: str,
@@ -270,7 +289,7 @@ def update_config_configmap(
     """
     # Check if source is a backup key (restore from memory)
     if source in _configmap_backups:
-        config_content = _configmap_backups[source]
+        config_content = _apply_inference_overrides(_configmap_backups[source])
         print(f"Restoring ConfigMap {configmap_name} from memory backup...")
 
         # Write content to temp file (oc create configmap requires a file)
@@ -289,12 +308,28 @@ def update_config_configmap(
                 os.remove(temp_path)
         return
 
-    # Otherwise, source is a file path
+    # Otherwise, source is a file path — apply inference overrides if needed
     print(f"Updating ConfigMap {configmap_name} with config from {source}...")
 
+    source_to_apply = source
+    temp_path = None
     try:
-        _recreate_configmap(configmap_name, source, configmap_key)
+        with open(source) as f:
+            patched = _apply_inference_overrides(f.read())
+        with open(source) as f:
+            original = f.read()
+        if patched != original:
+            tmp = tempfile.NamedTemporaryFile(mode="w", suffix=".yaml", delete=False)
+            tmp.write(patched)
+            tmp.close()
+            source_to_apply = tmp.name
+            temp_path = tmp.name
+
+        _recreate_configmap(configmap_name, source_to_apply, configmap_key)
         print(f"ConfigMap {configmap_name} updated successfully")
     except subprocess.CalledProcessError as e:
         print(f"Failed to update ConfigMap: {e}")
         raise
+    finally:
+        if temp_path and os.path.exists(temp_path):
+            os.remove(temp_path)