Feat/mobile backend - #7
Open
People62 wants to merge 28 commits into
Open
Conversation
- getMobileContext helper (Bearer userId + auto-resolved org) - GET/POST /api/mobile/assistants (list w/ user default id, create) - GET/PUT/DELETE /api/mobile/assistants/[id] - POST /api/mobile/assistants/[id]/default (per-user default agent) - POST /api/mobile/assistants/[id]/duplicate - GET /api/mobile/models (slim list for the builder dropdown)
Load the session's assistant and use its systemPrompt (persona), model, and modelConfig (temperature/topP/penalties). Clamp output tokens via the correct AI SDK v6 field maxOutputTokens (the old maxTokens was silently ignored).
Bearer-auth mobile routes over the existing (auth-agnostic) workflow service: - GET /api/mobile/workflows (org-scoped list) - GET/PUT(status)/DELETE /api/mobile/workflows/[id] - POST /api/mobile/workflows/[id]/execute (STANDARD only; CHATFLOW guarded) - GET /api/mobile/workflows/[id]/runs and /runs/[runId] authorizeMobileWorkflow enforces org scoping (web dashboard routes don't).
…xy, uploads) Bearer-auth mobile routes over the existing (auth-agnostic) media service: - POST/GET /api/mobile/media/jobs (generate image/audio, synchronous; VIDEO rejected) - GET /api/mobile/media/assets (org gallery) + /[id] GET/PATCH(favorite)/DELETE - GET /api/mobile/media/assets/[id]/file (proxy bytes so the phone never hits the internal S3 host — the key delivery fix) - GET /api/mobile/media/models?modality= - POST /api/mobile/media/uploads (multipart image → reference asset) Org context is required (media is org-scoped).
…WAV) OpenRouter requires stream:true for audio, and streaming only allows raw pcm16. Rather than change the shared web generateAudio, mobile audio uses its own path that streams pcm16 and wraps it into a playable WAV. The mobile jobs route branches AUDIO -> createMobileAudioJob, IMAGE -> shared createMediaJob. No web file is touched. Reuses the shared repository/storage (data layer) only.
React Native's <Image>/<Video> can't reliably attach an Authorization header on Android, so the /file proxy now also accepts the JWT via a ?token= query param (verified the same way). Adds userIdFromMobileToken helper.
Image models via OpenRouter ignore width/height (always square). Mobile now center-crops the generated image to the requested aspect ratio and persists the new bytes + dimensions (sharp). Mobile-only; web generation path unchanged.
Bearer-auth mobile routes over the existing (auth-decoupled, org-scoped) knowledge service: - GET/POST /api/mobile/knowledge/documents (list + upload; sync ingestion, maxDuration 600 since extract+OCR+chunk+embed run inline) - GET/PUT/DELETE /api/mobile/knowledge/documents/[id] - GET /api/mobile/knowledge/documents/[id]/file (byte proxy, ?token=) - GET/POST /api/mobile/knowledge/groups + /[id] (GET/PUT/DELETE) - GET/POST /api/mobile/knowledge/categories
…d ingest - GET /api/mobile/knowledge/documents/[id]/intelligence (entities + relations, org-authorized via the document) - upload defaults useEnhanced=true (entity/relation extraction) so the mobile Intelligence view has data; client can send enhanced=false to skip it
…es + tools/skills mgmt routes
Two levers against the ~30s full-scan cosine that dominates KB chat latency: - embeddings: send OpenAI-compatible `dimensions` = KB_EMBEDDING_DIM so query + re-embed can produce e.g. 1024-dim (qwen3 MRL) instead of 4096. Full-scan cost is ~O(N*dim), so 4096->1024 is a ~4x cut on its own. At dim=4096 (default) the body is identical to before (native), so the main corpus is unaffected. - hybrid vectorSearch: opt-in KB_VECTOR_KNN=true uses the SurrealDB HNSW KNN operator (<|K,ef|>) instead of full scan — sub-linear, sub-second once an HNSW index exists at the lower dim. Gated: only for UNSCOPED queries (the operator returns global K nearest and can't honor an extra WHERE filter), with a full-scan fallback on error, so scoped/multi-tenant paths are untouched. Both default OFF/native; enabled per-deployment via env. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GcfEvWTTaubz6wAhhaidHo
- bulk-re-embed: SurrealDB JS SDK v2 returns `id` from SELECT as a RecordId
object, so `UPDATE document_chunk:\`${c.id}\`` built a malformed target and
silently updated nothing — re-embeds reported success but never persisted.
Bind the RecordId as a param (`UPDATE $rid ...`). This is why the demo's
4096->1024 re-embed initially no-op'd.
- hybrid vectorSearch: KB_VECTOR_KNN now engages even for scoped queries (the
widget always scopes to its KB's document_ids). Over-fetch index candidates
(k=max(limit*5,100)) then AND the scope filter, so enough survive. Measured on
the 11.5k-chunk / 1024-dim demo corpus: full-scan 5452ms -> KNN 109ms unscoped,
616ms scoped. Still full-scan fallback on error; default OFF.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GcfEvWTTaubz6wAhhaidHo
…ARCH_ENABLED hybridRetrieve's entity arm runs an entity/graph traversal that, on a KB with no populated entity graph, returns 0 results yet adds large, variable latency (measured up to +28s on the 11.5k-chunk demo KB; entity results consistently 0). Default stays on; set KB_ENTITY_SEARCH_ENABLED=false to skip it — quality-neutral there and the single biggest remaining latency lever after HNSW KNN. Explicit options.enableEntitySearch still overrides. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GcfEvWTTaubz6wAhhaidHo
…equireMobileAdmin
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.