Architecture

SliceSeeker is a pnpm monorepo. Application code lives in workspace packages; production runs those packages as containers plus shared infrastructure.

Phase model

┌─────────────────────────────────────────────┐
│  Phase 1: Indexing                          │
│  admin-ui → indexer-api → Valkey queues     │
│                 ↓                           │
│           indexer-worker                    │
│                 ↓                           │
│         Postgres + S3                       │
└─────────────────────────────────────────────┘
                      │
                      ▼  shared schema
┌─────────────────────────────────────────────┐
│  Phase 2: Search                            │
│  search-api  ←→  Postgres                   │
│  (embeds the query, reads vectors only)     │
└─────────────────────────────────────────────┘

Phases do not call each other over HTTP. Phase 2 only needs a populated Postgres instance and an embedding provider for query vectors.

Application containers

Container	Source	Role
indexer-api	`server/`	Phase 1 HTTP API: uploads/hooks, collections, modality pipelines, search + costs used by the admin UI. Exposes `/health` and `/ready`.
indexer-worker	`worker/`	BullMQ consumers. Includes ffmpeg. Runs prep jobs (chunk / extract / sample) and API jobs (embed / ASR). Exits on startup if `AI_GATEWAY_API_KEY` is missing.
admin-ui	`client/`	Operator UI. Proxies API and tusd.
search-api	`search/`	Phase 2 read-only search: multimodal, transcript, frame, and hybrid RRF endpoints. `/ready` checks gateway key + schema.
db-push	`packages/db/`	One-shot: enable `vector` extension and push Drizzle schema.

Infrastructure containers

Service	Typical image	Role
db	`pgvector/pgvector`	Postgres + vectors
valkey	Valkey / Redis-compatible	BullMQ job state. Must use `maxmemory-policy noeviction`.
rustfs (self-contained layout)	S3-compatible	Object storage for source video, audio extracts, frames
tusd	`tusproject/tusd`	Resumable uploads to S3; HTTP hooks notify indexer-api

Shared packages

Package	Purpose
`db`	Drizzle schema and access layer (uploads, collections, chunks, transcripts, frames, costs)
`queue`	Job names, payloads, Valkey connection, concurrency helpers
`search-client`	Typed SDK for search-api

Queues

Phase 1 splits work across two BullMQ queues so ffmpeg-heavy prep can scale separately from embedding / Whisper load:

Queue	Kind of work	Concurrency env
Prep (`demo-prep-jobs`)	`chunk-video`, `extract-audio`, `sample-frames`, …	`PREP_WORKER_CONCURRENCY`
API (`demo-api-jobs`)	`embed-chunk`, `transcribe-part`, `embed-transcript`, `embed-frame`, …	`API_WORKER_CONCURRENCY`

See Operations for pool sizing, retries, and readiness checks.

Data flow (one modality)

Browser uploads via tusd → object lands in S3.
tusd hook → indexer-api records the upload.
Operator starts a pipeline → API enqueues a parent task.
indexer-worker runs prep then embed/ASR jobs.
Vectors and metadata land in Postgres.
search-api (or indexer-api from the admin UI) embeds the query and runs cosine similarity.