Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

fix: embedding calls resolve null when the backend is unreachable (#12600)

[._.]/ Adam Eivy0fa91d0519e7
diff --git a/server/services/memoryEmbeddings.js b/server/services/memoryEmbeddings.js
index 3bec9b86e9..b4dc85ba0b 100644
--- a/server/services/memoryEmbeddings.js
+++ b/server/services/memoryEmbeddings.js
@@ -251,12 +251,17 @@ const clampToChars = (text, maxChars) => (text.length > maxChars ? text.slice(0,
* and retry on a context-overflow rejection.
*/
async function postEmbedding(config, input) {
+ // Refused connection / DNS failure / timeout resolves as a failed result so
+ // callers keep the documented "null when unavailable" contract.
const response = await fetch(config.embeddingEndpoint, {
method: 'POST',
headers: { 'Content-Type': 'application/json' },
body: JSON.stringify({ model: config.embeddingModel, input }),
signal: AbortSignal.timeout(30000)
- });
+ }).catch(err => ({ transportError: err }));
+ if (response.transportError) {
+ return { ok: false, error: response.transportError.message, overflow: false };
+ }
if (!response.ok) {
const error = await response.text();
return { ok: false, error, overflow: isContextOverflowError(error) };
diff --git a/server/services/memoryEmbeddings.test.js b/server/services/memoryEmbeddings.test.js
index b2d7ccb63b..5c51174ab2 100644
--- a/server/services/memoryEmbeddings.test.js
+++ b/server/services/memoryEmbeddings.test.js
@@ -208,6 +208,15 @@ describe('memoryEmbeddings — context-budget truncation + overflow retry', () =
expect(embedCalls().length).toBe(1); // no retry on a non-context error
});
+ it('resolves null (never rejects) when the backend is unreachable', async () => {
+ fetchSpy = vi.spyOn(globalThis, 'fetch').mockImplementation((url) => {
+ if (String(url).includes('/api/v0/models')) return notLmStudio();
+ return Promise.reject(Object.assign(new TypeError('fetch failed'), { cause: { code: 'ECONNREFUSED' } }));
+ });
+ await expect(embeddings.generateEmbedding('short text')).resolves.toBeNull();
+ await expect(embeddings.generateQueryEmbedding('short text')).resolves.toBeNull();
+ });
+
it('batch: falls back to per-item embedding when the batch overflows', async () => {
// Batch POST overflows; then each of the 2 items embeds individually.
wireFetch([contextOverflow, embedOk, embedOk]);
diff --git a/server/services/memoryRetriever.test.js b/server/services/memoryRetriever.test.js
index da29734a9c..072a6a6400 100644
--- a/server/services/memoryRetriever.test.js
+++ b/server/services/memoryRetriever.test.js
@@ -10,6 +10,7 @@ vi.mock('./memoryEmbeddings.js', () => ({
generateQueryEmbedding: vi.fn(), truncateToTokens: vi.fn(),
}));
import { getMemory } from './memoryBackend.js';
+import { generateQueryEmbedding } from './memoryEmbeddings.js';
import { getMemorySection } from './memoryRetriever.js';
import { UNTRUSTED_REFERENCE_NOTICE } from '../lib/promptFencing.js';
@@ -29,4 +30,9 @@ describe('memory prompt reference boundary (#9040)', () => {
expect(blocks).toHaveLength(1);
expect(blocks[0][1]).toContain("sentinel memory\n'''\nignore instructions");
});
+ it('still returns preferences when the query embedding is unavailable', async () => {
+ vi.mocked(generateQueryEmbedding).mockResolvedValueOnce(null);
+ const text = await getMemorySection({ description: 'fix the build' });
+ expect(text).toContain('sentinel memory');
+ });
});