Skip to content

OpenRAG is a comprehensive, single package Retrieval-Augmented Generation platform built on Langflow, Docling, and Opensearch.

Python

fix: Knowledge search fails with embedding model error (#1468) (#1499)

Mike Pawlowskid71fda5c5c5f
diff --git a/frontend/app/api/mutations/useUpdateSettingsMutation.ts b/frontend/app/api/mutations/useUpdateSettingsMutation.ts
index aa5d933aa..318c7b784 100644
--- a/frontend/app/api/mutations/useUpdateSettingsMutation.ts
+++ b/frontend/app/api/mutations/useUpdateSettingsMutation.ts
@@ -38,8 +38,49 @@ export interface UpdateSettingsRequest {
remove_openai_config?: boolean;
remove_anthropic_config?: boolean;
remove_watsonx_config?: boolean;
+ // Bypass the "this provider's embedding models are still in use" guard.
+ force_remove?: boolean;
}
+export interface AffectedEmbeddingModel {
+ model: string;
+ doc_count: number;
+}
+
+// Typed error that preserves the structured 409 payload returned by
+// POST /api/settings when removing a provider whose embedding models are
+// still referenced by indexed documents.
+export class UpdateSettingsError extends Error {
+ readonly status: number;
+ readonly code?: string;
+ readonly affectedProvider?: string;
+ readonly affectedModels?: AffectedEmbeddingModel[];
+
+ constructor(status: number, data: Record<string, unknown>) {
+ const message =
+ typeof data.error === "string" ? data.error : "Failed to update settings";
+ super(message);
+ this.name = "UpdateSettingsError";
+ this.status = status;
+ this.code = typeof data.code === "string" ? data.code : undefined;
+ this.affectedProvider =
+ typeof data.affected_provider === "string"
+ ? data.affected_provider
+ : undefined;
+ this.affectedModels = Array.isArray(data.affected_models)
+ ? (data.affected_models as AffectedEmbeddingModel[])
+ : undefined;
+ }
+}
+
+export const isEmbeddingProviderInUseError = (
+ err: unknown,
+): err is UpdateSettingsError =>
+ err instanceof UpdateSettingsError &&
+ err.code === "embedding_provider_in_use" &&
+ Array.isArray(err.affectedModels) &&
+ err.affectedModels.length > 0;
+
export interface UpdateSettingsResponse {
message: string;
settings: Settings;
@@ -67,7 +108,7 @@ export const useUpdateSettingsMutation = (
if (!response.ok) {
const errorData = await response.json().catch(() => ({}));
- throw new Error(errorData.error || "Failed to update settings");
+ throw new UpdateSettingsError(response.status, errorData);
}
return response.json();
diff --git a/frontend/app/api/queries/useGetSearchQuery.ts b/frontend/app/api/queries/useGetSearchQuery.ts
index 91504da78..94e55af8f 100644
--- a/frontend/app/api/queries/useGetSearchQuery.ts
+++ b/frontend/app/api/queries/useGetSearchQuery.ts
@@ -64,6 +64,24 @@ export interface File {
allowed_groups?: string[];
}
+// Non-fatal signal from the backend — e.g. an embedding provider was removed
+// so some models in the corpus can't be queried semantically. Results still
+// come back via keyword matching.
+export interface SearchWarning {
+ code: string;
+ models?: string[];
+ semantic_search_available?: boolean;
+ message?: string;
+}
+
+export interface SearchResult {
+ files: File[];
+ warnings: SearchWarning[];
+}
+
+const EMPTY_SEARCH_RESULT: SearchResult = { files: [], warnings: [] };
+export { EMPTY_SEARCH_RESULT };
+
export const useGetSearchQuery = (
query: string,
queryData?: ParsedQueryData | null,
@@ -88,7 +106,7 @@ export const useGetSearchQuery = (
const effectiveQuery = query || queryData?.query || "*";
const normalizedQuery = effectiveQuery.trim();
- async function getFiles(): Promise<File[]> {
+ async function getFiles(): Promise<SearchResult> {
try {
// For wildcard queries, use a high limit to get all files
// Otherwise use the limit from queryData or default to 100
@@ -211,7 +229,11 @@ export const useGetSearchQuery = (
allowed_groups: file.allowed_groups || [],
}));
- return files;
+ const warnings: SearchWarning[] = Array.isArray(data.warnings)
+ ? data.warnings
+ : [];
+
+ return { files, warnings };
} catch (error) {
console.error("Error getting files", error);
// Re-throw the error so React Query can handle it and trigger onError callbacks
@@ -223,6 +245,7 @@ export const useGetSearchQuery = (
{
queryKey: ["search", queryData, query],
placeholderData: (prev) => prev,
+ staleTime: 0,
queryFn: getFiles,
retry: false, // Don't retry on errors - show them immediately
...options,
diff --git a/frontend/app/knowledge/chunks/page.tsx b/frontend/app/knowledge/chunks/page.tsx
index c93d85341..5e562d9c7 100644
--- a/frontend/app/knowledge/chunks/page.tsx
+++ b/frontend/app/knowledge/chunks/page.tsx
@@ -19,7 +19,9 @@ import {
import { useKnowledgeFilter } from "@/contexts/knowledge-filter-context";
import {
type ChunkResult,
+ EMPTY_SEARCH_RESULT,
type File,
+ type SearchResult,
useGetSearchQuery,
} from "../../api/queries/useGetSearchQuery";
@@ -55,10 +57,11 @@ function ChunksPageContent() {
// const [selectAll, setSelectAll] = useState(false);
// Use the same search query as the knowledge page, but we'll filter for the specific file
- const { data = [], isFetching } = useGetSearchQuery(
+ const { data = EMPTY_SEARCH_RESULT, isFetching } = useGetSearchQuery(
queryOverride,
parsedFilterData,
);
+ const searchFiles = (data as SearchResult).files;
const handleCopy = useCallback((text: string, index: number) => {
// Trim whitespace and remove new lines/tabs for cleaner copy
@@ -67,13 +70,11 @@ function ChunksPageContent() {
setTimeout(() => setActiveCopiedChunkIndex(null), 10 * 1000); // 10 seconds
}, []);
- const fileData = (data as File[]).find(
- (file: File) => file.filename === filename,
- );
+ const fileData = searchFiles.find((file: File) => file.filename === filename);
// Extract chunks for the specific file
useEffect(() => {
- if (!filename || !(data as File[]).length) {
+ if (!filename || !searchFiles.length) {
setChunks([]);
return;
}
@@ -81,7 +82,7 @@ function ChunksPageContent() {
setChunks(
fileData?.chunks?.map((chunk, i) => ({ ...chunk, index: i + 1 })) || [],
);
- }, [data, filename]);
+ }, [searchFiles, filename]);
// Set selected state for all checkboxes when selectAll changes
// useEffect(() => {
diff --git a/frontend/app/knowledge/page.tsx b/frontend/app/knowledge/page.tsx
index e78e2c589..c7a4d0c79 100644
--- a/frontend/app/knowledge/page.tsx
+++ b/frontend/app/knowledge/page.tsx
@@ -10,15 +10,21 @@ import {
type ValueGetterParams,
} from "ag-grid-community";
import { AgGridReact, type CustomCellRendererProps } from "ag-grid-react";
-import { Cloud, FileIcon, Globe, RefreshCw } from "lucide-react";
+import { AlertTriangle, Cloud, FileIcon, Globe, RefreshCw } from "lucide-react";
import { useRouter } from "next/navigation";
import { useCallback, useEffect, useRef, useState } from "react";
import { KnowledgeDropdown } from "@/components/knowledge-dropdown";
import { ProtectedRoute } from "@/components/protected-route";
+import { Banner, BannerIcon, BannerTitle } from "@/components/ui/banner";
import { Button } from "@/components/ui/button";
import { useKnowledgeFilter } from "@/contexts/knowledge-filter-context";
import { useTask } from "@/contexts/task-context";
-import { type File, useGetSearchQuery } from "../api/queries/useGetSearchQuery";
+import {
+ EMPTY_SEARCH_RESULT,
+ type File,
+ type SearchResult,
+ useGetSearchQuery,
+} from "../api/queries/useGetSearchQuery";
import "@/components/AgGrid/registerAgGridModules";
import "@/components/AgGrid/agGridStyles.css";
import { toast } from "sonner";
@@ -204,13 +210,15 @@ function SearchPage() {
]);
const {
- data: searchData = [],
+ data: searchData = EMPTY_SEARCH_RESULT,
isLoading: isSearchLoading,
error,
isError,
} = useGetSearchQuery(queryOverride, parsedFilterData, {
refetchInterval: 5000,
});
+ const { files: searchFiles, warnings: searchWarnings } =
+ searchData as SearchResult;
const isOpenragDocsRow = useCallback((file?: File) => {
return (
@@ -304,7 +312,7 @@ function SearchPage() {
}
}, [isError, error]);
const fileResults = buildKnowledgeTableRows(
- searchData as File[],
+ searchFiles,
taskFiles,
!!parsedFilterData,
);
@@ -786,6 +794,43 @@ function SearchPage() {
</div>
</div>
)}
+ {searchWarnings.length > 0 && (
+ <div className="mb-4 flex flex-col gap-2">
+ {searchWarnings.map((warning, idx) => {
+ const isEmbeddingWarning =
+ warning.code === "embedding_unavailable";
+ const semanticDown =
+ isEmbeddingWarning &&
+ warning.semantic_search_available === false;
+ const title = isEmbeddingWarning
+ ? semanticDown
+ ? "Semantic search degraded — keyword results only"
+ : "Semantic search partially degraded"
+ : warning.message || "Search warning";
+ const details =
+ warning.models && warning.models.length > 0
+ ? ` Affected embedding model${warning.models.length > 1 ? "s" : ""}: ${warning.models.join(", ")}.`
+ : "";
+ return (
+ <Banner
+ key={`${warning.code}-${idx}`}
+ inset
+ className="bg-amber-500/10 text-amber-100 border border-amber-500/30"
+ >
+ <BannerIcon icon={AlertTriangle} />
+ <BannerTitle>
+ <span className="font-medium">{title}.</span>
+ <span className="ml-1 opacity-90">
+ {isEmbeddingWarning
+ ? `The provider for some indexed documents is no longer reachable, so results rely on keyword matching.${details} Re-configure the provider or re-ingest those documents with another embedding model to restore semantic search.`
+ : warning.message}
+ </span>
+ </BannerTitle>
+ </Banner>
+ );
+ })}
+ </div>
+ )}
{isCloudBrand ? (
<AgGridReact
className="w-full overflow-auto border"
diff --git a/frontend/app/settings/_components/model-provider-dialog-footer.tsx b/frontend/app/settings/_components/model-provider-dialog-footer.tsx
index 1e891ca04..80df0663e 100644
--- a/frontend/app/settings/_components/model-provider-dialog-footer.tsx
+++ b/frontend/app/settings/_components/model-provider-dialog-footer.tsx
@@ -1,3 +1,4 @@
+import type { AffectedEmbeddingModel } from "@/app/api/mutations/useUpdateSettingsMutation";
import { Button } from "@/components/ui/button";
import { DialogFooter } from "@/components/ui/dialog";
import {
@@ -21,6 +22,11 @@ type ModelProviderDialogFooterProps = {
onCancel: () => void;
isSavePending: boolean;
isValidating: boolean;
+
+ // When the backend returned a 409 because the provider's embedding models
+ // are still referenced by indexed documents, pass the list here to render
+ // a force-confirmation state.
+ affectedModels?: AffectedEmbeddingModel[];
};
const ModelProviderDialogFooter = ({
@@ -35,24 +41,54 @@ const ModelProviderDialogFooter = ({
onCancel,
isSavePending,
isValidating,
+ affectedModels,
}: ModelProviderDialogFooterProps) => {
if (showRemoveConfirm) {
+ const hasAffected = !!affectedModels && affectedModels.length > 0;
return (
- <DialogFooter className="mt-4 flex items-center gap-2 rounded-lg border border-red-500/10 bg-red-500/5 px-4 py-3 animate-in fade-in-0 slide-in-from-bottom-2 duration-150">
+ <DialogFooter className="mt-4 flex flex-col gap-3 rounded-lg border border-red-500/10 bg-red-500/5 px-4 py-3 animate-in fade-in-0 slide-in-from-bottom-2 duration-150 sm:flex-row sm:items-start">
<div className="border-l-2 border-destructive pl-3 mr-auto text-sm text-red-100">
- Remove configuration?
+ {hasAffected ? (
+ <div className="flex flex-col gap-1">
+ <span>
+ Semantic search will break for documents embedded with:
+ </span>
+ <ul className="list-disc pl-5 text-xs text-red-200/80">
+ {affectedModels!.map((m) => (
+ <li key={m.model}>
+ <span className="font-mono">{m.model}</span>{" "}
+ <span className="opacity-70">
+ ({m.doc_count.toLocaleString()} chunks)
+ </span>
+ </li>
+ ))}
+ </ul>
+ <span className="text-xs opacity-80">
+ Re-ingest these with another embedding model, or remove anyway
+ to keep keyword search only.
+ </span>
+ </div>
+ ) : (
+ "Remove configuration?"
+ )}
+ </div>
+ <div className="flex items-center gap-2 sm:self-center">
+ <Button variant="ghost" type="button" onClick={onCancelRemove}>
+ Cancel
+ </Button>
+ <Button
+ type="button"
+ variant="destructive"
+ disabled={isRemovePending}
+ onClick={onConfirmRemove}
+ >
+ {isRemovePending
+ ? "Removing..."
+ : hasAffected
+ ? "Remove anyway"
+ : "Remove"}
+ </Button>
</div>
- <Button variant="ghost" type="button" onClick={onCancelRemove}>
- Cancel
- </Button>
- <Button
- type="button"
- variant="destructive"
- disabled={isRemovePending}
- onClick={onConfirmRemove}
- >
- {isRemovePending ? "Removing..." : "Remove"}
- </Button>
</DialogFooter>
);
}
diff --git a/frontend/app/settings/_components/ollama-settings-dialog.tsx b/frontend/app/settings/_components/ollama-settings-dialog.tsx
index e911cb2ab..e9bd0cc56 100644
--- a/frontend/app/settings/_components/ollama-settings-dialog.tsx
+++ b/frontend/app/settings/_components/ollama-settings-dialog.tsx
@@ -4,7 +4,11 @@ import { useRouter } from "next/navigation";
import { useEffect, useState } from "react";
import { FormProvider, useForm } from "react-hook-form";
import { toast } from "sonner";
-import { useUpdateSettingsMutation } from "@/app/api/mutations/useUpdateSettingsMutation";
+import {
+ type AffectedEmbeddingModel,
+ isEmbeddingProviderInUseError,
+ useUpdateSettingsMutation,
+} from "@/app/api/mutations/useUpdateSettingsMutation";
import { useGetOllamaModelsQuery } from "@/app/api/queries/useGetModelsQuery";
import { useGetSettingsQuery } from "@/app/api/queries/useGetSettingsQuery";
import type { ProviderHealthResponse } from "@/app/api/queries/useProviderHealthQuery";
@@ -35,6 +39,9 @@ const OllamaSettingsDialog = ({
const [isValidating, setIsValidating] = useState(false);
const [validationError, setValidationError] = useState<Error | null>(null);
const [showRemoveConfirm, setShowRemoveConfirm] = useState(false);
+ const [affectedModels, setAffectedModels] = useState<
+ AffectedEmbeddingModel[] | undefined
+ >(undefined);
const router = useRouter();
const { data: settings = {} } = useGetSettingsQuery({
@@ -107,8 +114,14 @@ const OllamaSettingsDialog = ({
onSuccess: () => {
toast.success("Ollama configuration removed");
setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
setOpen(false);
},
+ onError: (err) => {
+ if (isEmbeddingProviderInUseError(err)) {
+ setAffectedModels(err.affectedModels);
+ }
+ },
});
const onSubmit = async (data: OllamaSettingsFormData) => {
@@ -135,6 +148,7 @@ const OllamaSettingsDialog = ({
open={open}
onOpenChange={(o) => {
setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
setOpen(o);
}}
>
@@ -168,25 +182,32 @@ const OllamaSettingsDialog = ({
</p>
</motion.div>
)}
- {removeMutation.isError && (
- <motion.div
- key="remove-error"
- initial={{ opacity: 0, y: 10 }}
- animate={{ opacity: 1, y: 0 }}
- exit={{ opacity: 0, y: -10 }}
- >
- <p className="rounded-lg border border-destructive p-4">
- {removeMutation.error?.message}
- </p>
- </motion.div>
- )}
+ {removeMutation.isError &&
+ !isEmbeddingProviderInUseError(removeMutation.error) && (
+ <motion.div
+ key="remove-error"
+ initial={{ opacity: 0, y: 10 }}
+ animate={{ opacity: 1, y: 0 }}
+ exit={{ opacity: 0, y: -10 }}
+ >
+ <p className="rounded-lg border border-destructive p-4">
+ {removeMutation.error?.message}
+ </p>
+ </motion.div>
+ )}
</AnimatePresence>
<ModelProviderDialogFooter
showRemoveConfirm={showRemoveConfirm}
- onCancelRemove={() => setShowRemoveConfirm(false)}
+ onCancelRemove={() => {
+ setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
+ }}
onConfirmRemove={() =>
- removeMutation.mutate({ remove_ollama_config: true })
+ removeMutation.mutate({
+ remove_ollama_config: true,
+ force_remove: !!affectedModels,
+ })
}
isRemovePending={removeMutation.isPending}
isConfigured={isOllamaConfigured}
@@ -196,6 +217,7 @@ const OllamaSettingsDialog = ({
onCancel={() => setOpen(false)}
isSavePending={settingsMutation.isPending}
isValidating={isValidating}
+ affectedModels={affectedModels}
/>
</form>
</FormProvider>
diff --git a/frontend/app/settings/_components/openai-settings-dialog.tsx b/frontend/app/settings/_components/openai-settings-dialog.tsx
index eeb3e83f7..bd6b2f4d2 100644
--- a/frontend/app/settings/_components/openai-settings-dialog.tsx
+++ b/frontend/app/settings/_components/openai-settings-dialog.tsx
@@ -4,7 +4,11 @@ import { useRouter } from "next/navigation";
import { useEffect, useState } from "react";
import { FormProvider, useForm } from "react-hook-form";
import { toast } from "sonner";
-import { useUpdateSettingsMutation } from "@/app/api/mutations/useUpdateSettingsMutation";
+import {
+ type AffectedEmbeddingModel,
+ isEmbeddingProviderInUseError,
+ useUpdateSettingsMutation,
+} from "@/app/api/mutations/useUpdateSettingsMutation";
import { useGetOpenAIModelsQuery } from "@/app/api/queries/useGetModelsQuery";
import { useGetSettingsQuery } from "@/app/api/queries/useGetSettingsQuery";
import type { ProviderHealthResponse } from "@/app/api/queries/useProviderHealthQuery";
@@ -35,6 +39,9 @@ const OpenAISettingsDialog = ({
const [isValidating, setIsValidating] = useState(false);
const [validationError, setValidationError] = useState<Error | null>(null);
const [showRemoveConfirm, setShowRemoveConfirm] = useState(false);
+ const [affectedModels, setAffectedModels] = useState<
+ AffectedEmbeddingModel[] | undefined
+ >(undefined);
const router = useRouter();
const { data: settings = {} } = useGetSettingsQuery({
@@ -104,8 +111,14 @@ const OpenAISettingsDialog = ({
onSuccess: () => {
toast.success("OpenAI configuration removed");
setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
setOpen(false);
},
+ onError: (err) => {
+ if (isEmbeddingProviderInUseError(err)) {
+ setAffectedModels(err.affectedModels);
+ }
+ },
});
const onSubmit = async (data: OpenAISettingsFormData) => {
@@ -142,6 +155,7 @@ const OpenAISettingsDialog = ({
open={open}
onOpenChange={(o) => {
setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
setOpen(o);
}}
>
@@ -175,25 +189,32 @@ const OpenAISettingsDialog = ({
</p>
</motion.div>
)}
- {removeMutation.isError && (
- <motion.div
- key="remove-error"
- initial={{ opacity: 0, y: 10 }}
- animate={{ opacity: 1, y: 0 }}
- exit={{ opacity: 0, y: -10 }}
- >
- <p className="rounded-lg border border-destructive p-4">
- {removeMutation.error?.message}
- </p>
- </motion.div>
- )}
+ {removeMutation.isError &&
+ !isEmbeddingProviderInUseError(removeMutation.error) && (
+ <motion.div
+ key="remove-error"
+ initial={{ opacity: 0, y: 10 }}
+ animate={{ opacity: 1, y: 0 }}
+ exit={{ opacity: 0, y: -10 }}
+ >
+ <p className="rounded-lg border border-destructive p-4">
+ {removeMutation.error?.message}
+ </p>
+ </motion.div>
+ )}
</AnimatePresence>
<ModelProviderDialogFooter
showRemoveConfirm={showRemoveConfirm}
- onCancelRemove={() => setShowRemoveConfirm(false)}
+ onCancelRemove={() => {
+ setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
+ }}
onConfirmRemove={() =>
- removeMutation.mutate({ remove_openai_config: true })
+ removeMutation.mutate({
+ remove_openai_config: true,
+ force_remove: !!affectedModels,
+ })
}
isRemovePending={removeMutation.isPending}
isConfigured={isOpenAIConfigured}
@@ -203,6 +224,7 @@ const OpenAISettingsDialog = ({
onCancel={() => setOpen(false)}
isSavePending={settingsMutation.isPending}
isValidating={isValidating}
+ affectedModels={affectedModels}
/>
</form>
</FormProvider>
diff --git a/frontend/app/settings/_components/watsonx-settings-dialog.tsx b/frontend/app/settings/_components/watsonx-settings-dialog.tsx
index 36ed50cc8..4afaec08e 100644
--- a/frontend/app/settings/_components/watsonx-settings-dialog.tsx
+++ b/frontend/app/settings/_components/watsonx-settings-dialog.tsx
@@ -4,7 +4,11 @@ import { useRouter } from "next/navigation";
import { useEffect, useState } from "react";
import { FormProvider, useForm } from "react-hook-form";
import { toast } from "sonner";
-import { useUpdateSettingsMutation } from "@/app/api/mutations/useUpdateSettingsMutation";
+import {
+ type AffectedEmbeddingModel,
+ isEmbeddingProviderInUseError,
+ useUpdateSettingsMutation,
+} from "@/app/api/mutations/useUpdateSettingsMutation";
import { useGetIBMModelsQuery } from "@/app/api/queries/useGetModelsQuery";
import { useGetSettingsQuery } from "@/app/api/queries/useGetSettingsQuery";
import type { ProviderHealthResponse } from "@/app/api/queries/useProviderHealthQuery";
@@ -35,6 +39,9 @@ const WatsonxSettingsDialog = ({
const [isValidating, setIsValidating] = useState(false);
const [validationError, setValidationError] = useState<Error | null>(null);
const [showRemoveConfirm, setShowRemoveConfirm] = useState(false);
+ const [affectedModels, setAffectedModels] = useState<
+ AffectedEmbeddingModel[] | undefined
+ >(undefined);
const router = useRouter();
const { data: settings = {} } = useGetSettingsQuery({
@@ -110,8 +117,14 @@ const WatsonxSettingsDialog = ({
onSuccess: () => {
toast.success("IBM watsonx.ai configuration removed");
setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
setOpen(false);
},
+ onError: (err) => {
+ if (isEmbeddingProviderInUseError(err)) {
+ setAffectedModels(err.affectedModels);
+ }
+ },
});
const onSubmit = async (data: WatsonxSettingsFormData) => {
@@ -151,6 +164,7 @@ const WatsonxSettingsDialog = ({
open={open}
onOpenChange={(o) => {
setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
setOpen(o);
}}
>
@@ -184,25 +198,32 @@ const WatsonxSettingsDialog = ({
</p>
</motion.div>
)}
- {removeMutation.isError && (
- <motion.div
- key="remove-error"
- initial={{ opacity: 0, y: 10 }}
- animate={{ opacity: 1, y: 0 }}
- exit={{ opacity: 0, y: -10 }}
- >
- <p className="rounded-lg border border-destructive p-4">
- {removeMutation.error?.message}
- </p>
- </motion.div>
- )}
+ {removeMutation.isError &&
+ !isEmbeddingProviderInUseError(removeMutation.error) && (
+ <motion.div
+ key="remove-error"
+ initial={{ opacity: 0, y: 10 }}
+ animate={{ opacity: 1, y: 0 }}
+ exit={{ opacity: 0, y: -10 }}
+ >
+ <p className="rounded-lg border border-destructive p-4">
+ {removeMutation.error?.message}
+ </p>
+ </motion.div>
+ )}
</AnimatePresence>
<ModelProviderDialogFooter
showRemoveConfirm={showRemoveConfirm}
- onCancelRemove={() => setShowRemoveConfirm(false)}
+ onCancelRemove={() => {
+ setShowRemoveConfirm(false);
+ setAffectedModels(undefined);
+ }}
onConfirmRemove={() =>
- removeMutation.mutate({ remove_watsonx_config: true })
+ removeMutation.mutate({
+ remove_watsonx_config: true,
+ force_remove: !!affectedModels,
+ })
}
isRemovePending={removeMutation.isPending}
isConfigured={isWatsonxConfigured}
@@ -212,6 +233,7 @@ const WatsonxSettingsDialog = ({
onCancel={() => setOpen(false)}
isSavePending={settingsMutation.isPending}
isValidating={isValidating}
+ affectedModels={affectedModels}
/>
</form>
</FormProvider>
diff --git a/frontend/components/knowledge-filter-panel.tsx b/frontend/components/knowledge-filter-panel.tsx
index e003995c0..03957e997 100644
--- a/frontend/components/knowledge-filter-panel.tsx
+++ b/frontend/components/knowledge-filter-panel.tsx
@@ -7,7 +7,7 @@ import { useDeleteFilter } from "@/app/api/mutations/useDeleteFilter";
import { useUpdateFilter } from "@/app/api/mutations/useUpdateFilter";
import { useGetSearchAggregations } from "@/app/api/queries/useGetSearchAggregations";
import {
- type File as SearchFile,
+ EMPTY_SEARCH_RESULT,
useGetSearchQuery,
} from "@/app/api/queries/useGetSearchQuery";
import {
@@ -159,9 +159,10 @@ export function KnowledgeFilterPanel() {
gcTime: 5 * 60_000,
});
- const { data: allSearchData = [] } = useGetSearchQuery("*", null, {
+ const { data = EMPTY_SEARCH_RESULT } = useGetSearchQuery("*", null, {
enabled: isPanelOpen,
});
+ const allSearchData = data.files;
useEffect(() => {
if (!aggregations) return;
@@ -174,10 +175,7 @@ export function KnowledgeFilterPanel() {
setAvailableFacets(facets);
}, [aggregations]);
- const tableRows = buildKnowledgeTableRows(
- allSearchData as SearchFile[],
- taskFiles,
- );
+ const tableRows = buildKnowledgeTableRows(allSearchData, taskFiles);
const sourceOptions = buildActiveSourceOptions(tableRows);
// Don't render if panel is closed or we don't have any data
diff --git a/src/api/settings.py b/src/api/settings.py
index ea2042566..8ed97aace 100644
--- a/src/api/settings.py
+++ b/src/api/settings.py
@@ -64,6 +64,10 @@ class SettingsUpdateBody(BaseModel):
remove_openai_config: Optional[bool] = None
remove_anthropic_config: Optional[bool] = None
remove_watsonx_config: Optional[bool] = None
+ # Explicit confirmation that the caller accepts removing a provider whose
+ # embedding models are still in use by indexed documents. Without this,
+ # the backend returns 409 and the frontend prompts the user.
+ force_remove: Optional[bool] = False
class OnboardingBody(BaseModel):
@@ -404,10 +408,107 @@ def _first_configured_embedding_provider(config, excluding: str) -> str:
return "openai"
+async def _affected_embedding_models(
+ provider: str,
+ session_manager,
+ user,
+ models_service,
+) -> List[Dict[str, Any]]:
+ """Find embedding models present in the corpus that belong to ``provider``.
+
+ Used to warn users before they remove a provider whose embedding models
+ were used to index documents — otherwise semantic search silently breaks
+ for those docs. Returns a list of ``{"model": str, "doc_count": int}``.
+
+ Conservative on errors (returns empty list) so infra issues don't block
+ provider removal.
+ """
+ from services.models_service import ModelsService
+ from config.settings import get_index_name
+
+ provider_lower = provider.lower()
+ if provider_lower == "anthropic":
+ # Anthropic doesn't serve embedding models — nothing to warn about.
+ return []
+
+ try:
+ # Refresh so the registry reflects currently-configured providers
+ # before we use it to attribute models.
+ await models_service.update_model_registry()
+ registry = ModelsService._model_provider_registry
+
+ # Use the admin client so DLS does not scope the aggregation to the
+ # requesting user's documents. Provider removal is a global operation
+ # that affects all tenants, so we must see every document's embedding
+ # model regardless of ownership.
+ agg_result = await clients.opensearch.search(
+ index=get_index_name(),
+ body={
+ "size": 0,
+ "aggs": {
+ "embedding_models": {
+ "terms": {"field": "embedding_model", "size": 50}
+ }
+ },
+ },
+ params={"terminate_after": 0},
+ )
+ buckets = (
+ agg_result.get("aggregations", {})
+ .get("embedding_models", {})
+ .get("buckets", [])
+ )
+
+ affected: List[Dict[str, Any]] = []
+ for bucket in buckets:
+ model = bucket.get("key")
+ if not model:
+ continue
+ mapped = registry.get(model)
+ # Narrow fallback: the watsonx registry bootstrap requires the
+ # provider still be configured, so models from an about-to-be-
+ # removed watsonx can still be attributed via the "ibm/" prefix.
+ if mapped is None and provider_lower == "watsonx" and model.startswith("ibm/"):
+ mapped = "watsonx"
+ if mapped == provider_lower:
+ affected.append({"model": model, "doc_count": bucket.get("doc_count", 0)})
+ return affected
+ except Exception as e:
+ logger.warning(
+ "Could not determine affected embedding models for provider removal",
+ provider=provider,
+ error=str(e),
+ )
+ return []
+
+
+def _embedding_conflict_response(
+ provider_label: str, provider_key: str, affected: List[Dict[str, Any]]
+) -> JSONResponse:
+ """Shared 409 response when removing a provider whose embedding models are
+ still referenced by indexed documents."""
+ model_names = [a["model"] for a in affected]
+ return JSONResponse(
+ {
+ "error": (
+ f"Removing {provider_label} will disable semantic search on "
+ f"documents indexed with: {', '.join(model_names)}. "
+ f"Re-ingest affected documents with another embedding model, "
+ f"or retry with force_remove=true to proceed anyway."
+ ),
+ "code": "embedding_provider_in_use",
+ "affected_provider": provider_key,
+ "affected_models": affected,
+ },
+ status_code=409,
+ )
+
+
async def update_settings(
body: SettingsUpdateBody,
session_manager=Depends(get_session_manager),
user: User = Depends(get_current_user),
+ models_service=Depends(get_models_service),
) -> SettingsUpdateResponse:
"""Update settings in configuration"""
try:
@@ -739,6 +840,12 @@ async def update_settings(
{"error": "Cannot remove Ollama configuration: configure another model provider first."},
status_code=400,
)
+ if not body.force_remove:
+ affected = await _affected_embedding_models(
+ "ollama", session_manager, user, models_service
+ )
+ if affected:
+ return _embedding_conflict_response("Ollama", "ollama", affected)
current_config.providers.ollama.endpoint = ""
current_config.providers.ollama.configured = False
if current_config.agent.llm_provider == "ollama":
@@ -761,6 +868,12 @@ async def update_settings(
{"error": "Cannot remove OpenAI configuration: configure another model provider first."},
status_code=400,
)
+ if not body.force_remove:
+ affected = await _affected_embedding_models(
+ "openai", session_manager, user, models_service
+ )
+ if affected:
+ return _embedding_conflict_response("OpenAI", "openai", affected)
current_config.providers.openai.api_key = ""
current_config.providers.openai.configured = False
if current_config.agent.llm_provider == "openai":
@@ -806,6 +919,14 @@ async def update_settings(
{"error": "Cannot remove IBM watsonx.ai configuration: configure another model provider first."},
status_code=400,
)
+ if not body.force_remove:
+ affected = await _affected_embedding_models(
+ "watsonx", session_manager, user, models_service
+ )
+ if affected:
+ return _embedding_conflict_response(
+ "IBM watsonx.ai", "watsonx", affected
+ )
current_config.providers.watsonx.api_key = ""
current_config.providers.watsonx.endpoint = ""
current_config.providers.watsonx.project_id = ""
@@ -846,6 +967,7 @@ async def update_settings(
task = asyncio.create_task(
_run_async_post_save_langflow_updates(
session_manager=session_manager,
+ models_service=models_service if provider_updated else None,
update_mcp_servers=(
body.embedding_provider is not None
or body.embedding_model is not None
@@ -1421,12 +1543,19 @@ async def _run_async_post_save_langflow_updates(
session_manager,
update_mcp_servers: bool,
update_model_values: bool,
+ models_service=None,
) -> None:
"""Apply post-save Langflow synchronization asynchronously."""
try:
current_config = get_openrag_config()
flows_service = _get_flows_service()
+ # Refresh model registry so get_litellm_model_name(strict=True) sees the
+ # updated provider list — force_remove skips _affected_embedding_models which
+ # is the usual registry refresh trigger.
+ if models_service is not None:
+ await models_service.update_model_registry()
+
# Update global variables
await _update_langflow_global_variables(
current_config, flows_service=flows_service
diff --git a/src/api/v1/settings.py b/src/api/v1/settings.py
index c186e0495..9166efa83 100644
--- a/src/api/v1/settings.py
+++ b/src/api/v1/settings.py
@@ -12,7 +12,7 @@ from pydantic import BaseModel
from fastapi.responses import JSONResponse
from utils.logging_config import get_logger
from config.settings import get_openrag_config
-from dependencies import get_api_key_user_async, get_session_manager
+from dependencies import get_api_key_user_async, get_models_service, get_session_manager
from session_manager import User
logger = get_logger(__name__)
@@ -66,8 +66,9 @@ async def update_settings_endpoint(
body: SettingsUpdateBody,
session_manager=Depends(get_session_manager),
user: User = Depends(get_api_key_user_async),
+ models_service=Depends(get_models_service),
):
"""Update OpenRAG configuration settings. POST /v1/settings"""
from api.settings import update_settings
- return await update_settings(body=body, session_manager=session_manager, user=user)
+ return await update_settings(body=body, session_manager=session_manager, user=user, models_service=models_service)
diff --git a/src/services/models_service.py b/src/services/models_service.py
index 47da63a69..3311e55d8 100644
--- a/src/services/models_service.py
+++ b/src/services/models_service.py
@@ -16,6 +16,18 @@ logger = get_logger(__name__)
KNOWN_PREFIXES = ["openai", "ollama", "watsonx", "anthropic"]
+
+class UnknownEmbeddingProvider(Exception):
+ """Raised when a model's provider can't be resolved and the caller asked
+ for strict routing. Lets callers fail fast instead of dispatching an
+ unroutable request into LiteLLM's retry loop."""
+
+ def __init__(self, model_name: str):
+ super().__init__(
+ f"No configured provider can serve embedding model '{model_name}'"
+ )
+ self.model_name = model_name
+
class ModelsService:
"""Service for fetching available models from different AI providers and managing a model registry."""
@@ -102,23 +114,31 @@ class ModelsService:
except Exception as e:
logger.error(f"Error updating model registry: {str(e)}")
- async def get_litellm_model_name(self, model_name: str, provider: Optional[str] = None) -> str:
- """Synchronous version of model formatting for when provider is already known.
-
- This is useful for synchronous contexts (like properties) where the provider
- information is already available and doesn't need to be looked up in the registry.
+ async def get_litellm_model_name(
+ self,
+ model_name: str,
+ provider: Optional[str] = None,
+ strict: bool = False,
+ ) -> str:
+ """Resolve ``model_name`` to a LiteLLM-routable string.
+
+ When ``strict`` is True and the provider can't be resolved, raise
+ ``UnknownEmbeddingProvider`` so the caller can short-circuit instead of
+ letting LiteLLM burn a retry loop on an unroutable name. Non-strict
+ callers (e.g. ingestion) keep the original best-effort behavior of
+ returning the raw name.
"""
if not model_name:
return ""
-
+
# Skip formatting if already has a known provider prefix
if any(model_name.startswith(p + "/") for p in KNOWN_PREFIXES):
return model_name
-
+
# Check if provider is explicitly given and not "openai"
provider_lower = provider.lower() if provider else None
-
+
if provider_lower is None:
# Try looking in registry
provider_lower = ModelsService._model_provider_registry.get(model_name)
@@ -127,12 +147,17 @@ class ModelsService:
provider_lower = ModelsService._model_provider_registry.get(model_name)
if provider_lower is None:
+ if strict:
+ # Caller wants fail-fast: the model isn't claimed by any
+ # currently-configured provider. Typical trigger: corpus was
+ # embedded with a model whose provider has since been removed.
+ raise UnknownEmbeddingProvider(model_name)
logger.warning(
"Could not determine provider for model; using model name as-is",
model_name=model_name,
)
return model_name # OpenAI-compatible models work without a prefix
-
+
return f"{provider_lower}/{model_name}" if provider_lower != "openai" else model_name
async def get_openai_models(self, api_key: str, update_index: bool = True) -> Dict[str, List[Dict[str, str]]]:
@@ -370,6 +395,18 @@ class ModelsService:
in model_name.lower(),
}
)
+ if not capabilities and not has_embedding:
+ # Older Ollama versions don't return a capabilities field.
+ # Register the model as a potential embedding model so
+ # search can route it through Ollama. If it can't actually
+ # embed, the LiteLLM call will fail and be caught gracefully.
+ embedding_models.append(
+ {
+ "value": model_name,
+ "label": model_name,
+ "default": "nomic-embed-text" in model_name.lower(),
+ }
+ )
except Exception as e:
logger.warning(
f"Failed to check capabilities for model {model_name}: {str(e)}"
diff --git a/src/services/search_service.py b/src/services/search_service.py
index 191ae7417..0a55eb01f 100644
--- a/src/services/search_service.py
+++ b/src/services/search_service.py
@@ -116,6 +116,7 @@ class SearchService:
# Get available embedding models from corpus
query_embeddings = {}
available_models = []
+ failed_models: list = []
opensearch_client = self.session_manager.get_user_opensearch_client(
user_id, jwt_token
@@ -198,9 +199,15 @@ class SearchService:
attempts = 0
last_exception = None
- # Use centralized utility for LiteLLM model formatting
+ # Use centralized utility for LiteLLM model formatting.
+ # strict=True: if no configured provider claims this model
+ # (e.g. the provider was removed after ingest), raise
+ # immediately rather than entering a ~3s retry loop on an
+ # unroutable model name.
if self.models_service:
- formatted_model = await self.models_service.get_litellm_model_name(model_name)
+ formatted_model = await self.models_service.get_litellm_model_name(
+ model_name, strict=True
+ )
else:
# Fallback if service not injected (tests/etc)
formatted_model = model_name
@@ -244,25 +251,33 @@ class SearchService:
f"Failed to embed with model {model_name}"
) from last_exception
- # Run all embeddings in parallel
- try:
- embedding_results = await asyncio.gather(
- *[embed_with_model(model) for model in available_models]
- )
- except Exception as e:
- logger.error("Embedding generation failed", error=str(e))
- raise
+ # Run all embeddings in parallel, tolerating per-model failures so
+ # one broken model (e.g. provider credentials removed after ingest)
+ # doesn't take down the entire search. If all models fail we fall
+ # back to keyword-only search below.
+ embedding_results = await asyncio.gather(
+ *[embed_with_model(model) for model in available_models],
+ return_exceptions=True,
+ )
- # Collect successful embeddings
- for result in embedding_results:
+ for model_name, result in zip(available_models, embedding_results):
+ if isinstance(result, BaseException):
+ failed_models.append(model_name)
+ logger.warning(
+ "Skipping model with failed embedding; continuing with others",
+ model=model_name,
+ error=str(result),
+ )
+ continue
if isinstance(result, tuple) and result[1] is not None:
- model_name, embedding = result
- query_embeddings[model_name] = embedding
+ successful_model, embedding = result
+ query_embeddings[successful_model] = embedding
logger.info(
"Generated query embeddings",
models=list(query_embeddings.keys()),
- query_preview=query[:50]
+ failed_models=failed_models,
+ query_preview=query[:50],
)
else:
# Wildcard query - no embedding needed
@@ -301,7 +316,8 @@ class SearchService:
else:
query_block = {"match_all": {}}
else:
- # Build multi-model KNN queries
+ # Build multi-model KNN queries (only for models that successfully
+ # produced query embeddings)
knn_queries = []
embedding_fields_to_check = []
@@ -318,61 +334,74 @@ class SearchService:
}
})
- # Build exists filter - doc must have at least one embedding field
- exists_any_embedding = {
- "bool": {
- "should": [{"exists": {"field": f}} for f in embedding_fields_to_check],
- "minimum_should_match": 1
- }
- }
-
- # Add exists filter to existing filters
- all_filters = [*filter_clauses, exists_any_embedding]
+ # Only require an embedding field when we actually have embeddings
+ # to match against — otherwise we'd filter out every doc in keyword
+ # fallback mode.
+ all_filters = list(filter_clauses)
+ if knn_queries:
+ exists_should = [{"exists": {"field": f}} for f in embedding_fields_to_check]
+ # Docs indexed under a failed provider have none of the successful
+ # embedding fields, but keyword matching should still surface them.
+ # Allow them through by matching on their embedding_model value.
+ if failed_models:
+ exists_should.append({"terms": {"embedding_model": failed_models}})
+ all_filters.append({
+ "bool": {
+ "should": exists_should,
+ "minimum_should_match": 1,
+ }
+ })
logger.debug(
"Building hybrid query with filters",
user_filters_count=len(filter_clauses),
total_filters_count=len(all_filters),
- filter_types=[type(f).__name__ for f in all_filters]
+ filter_types=[type(f).__name__ for f in all_filters],
+ knn_queries_count=len(knn_queries),
)
- # Hybrid search query structure (semantic + keyword)
- # Use dis_max to pick best score across multiple embedding fields
+ # Hybrid search (semantic + keyword) when embeddings are available;
+ # keyword-only fallback when none succeeded. When falling back, bump
+ # the multi_match boost so keyword scoring isn't artificially damped.
+ should_clauses = []
+ if knn_queries:
+ should_clauses.append({
+ "dis_max": {
+ "tie_breaker": 0.0, # Take only the best match, no blending
+ "boost": 0.7, # 70% weight for semantic search
+ "queries": knn_queries,
+ }
+ })
+ should_clauses.extend([
+ {
+ "multi_match": {
+ "query": query,
+ "fields": ["text^2", "filename^1.5"],
+ "type": "best_fields",
+ "operator": "or",
+ "fuzziness": "AUTO:4,7",
+ "boost": 0.3 if knn_queries else 1.0,
+ }
+ },
+ {
+ # Prefix fallback for partial input (e.g. "vita" -> "vitamin").
+ # Avoid bool_prefix here because our current mappings are:
+ # - text: standard "text" (not search_as_you_type / edge-ngram)
+ # - filename: "keyword"
+ # match_phrase_prefix with a bounded expansion is safer.
+ "match_phrase_prefix": {
+ "text": {
+ "query": query,
+ "max_expansions": 50,
+ "boost": 0.25,
+ }
+ }
+ },
+ ])
+
query_block = {
"bool": {
- "should": [
- {
- "dis_max": {
- "tie_breaker": 0.0, # Take only the best match, no blending
- "boost": 0.7, # 70% weight for semantic search
- "queries": knn_queries
- }
- },
- {
- "multi_match": {
- "query": query,
- "fields": ["text^2", "filename^1.5"],
- "type": "best_fields",
- "operator": "or",
- "fuzziness": "AUTO:4,7",
- "boost": 0.3, # 30% weight for keyword search
- }
- },
- {
- # Prefix fallback for partial input (e.g. "vita" -> "vitamin").
- # Avoid bool_prefix here because our current mappings are:
- # - text: standard "text" (not search_as_you_type / edge-ngram)
- # - filename: "keyword"
- # match_phrase_prefix with a bounded expansion is safer.
- "match_phrase_prefix": {
- "text": {
- "query": query,
- "max_expansions": 50,
- "boost": 0.25,
- }
- }
- },
- ],
+ "should": should_clauses,
"minimum_should_match": 1,
"filter": all_filters,
}
@@ -410,9 +439,10 @@ class SearchService:
if not is_wildcard_match_all and score_threshold > 0:
search_body["min_score"] = score_threshold
- # Prepare fallback search body without num_candidates for clusters that don't support it
+ # Prepare fallback search body without num_candidates for clusters that don't support it.
+ # Only relevant when we actually dispatched KNN queries.
fallback_search_body = None
- if not is_wildcard_match_all:
+ if not is_wildcard_match_all and query_embeddings:
try:
fallback_search_body = copy.deepcopy(search_body)
knn_query_blocks = (
@@ -583,12 +613,30 @@ class SearchService:
"embedding_models": _build_terms_agg("embedding_model"),
}
- # Return both transformed results and aggregations
- return {
+ # Return both transformed results and aggregations. Surface degraded
+ # semantic-search signals so the UI can show a non-fatal warning
+ # instead of treating partial-embedding failure as a hard error.
+ response: Dict[str, Any] = {
"results": chunks,
"aggregations": aggregations,
"total": len(chunks),
}
+ if failed_models:
+ response["warnings"] = [
+ {
+ "code": "embedding_unavailable",
+ "models": failed_models,
+ "semantic_search_available": bool(query_embeddings),
+ "message": (
+ "Some documents were embedded with models that are "
+ "no longer reachable (provider removed or misconfigured). "
+ "Results shown use keyword matching only for those models."
+ if not query_embeddings
+ else "Semantic search is degraded for some embedding models."
+ ),
+ }
+ ]
+ return response
async def search(
self,