Cross-primitive Recipes

These recipes span multiple K3 primitives — Storage + Pipelines + Vector + Tables — in a single workflow. They’re the platform-level story: how one bucket becomes a knowledge base that’s queryable as bytes, embeddings, AND structured rows simultaneously.

For recipes scoped to a single primitive (Storage’s browser upload, Vector’s hybrid+rerank, Tables’ time-travel, etc.) see the per-primitive recipe folders.

What’s here

Recipe	Primitives involved	When you need it
RAG Knowledge Base	Storage · Pipelines · Vector	Drop PDFs in a bucket → search them semantically. The canonical “Hello, K3” full-stack scenario.
Mixed-Media Library	Storage · Pipelines · Vector	Upload images / video / audio → search by visual similarity or by file. Showcases multimodal end-to-end.
Reviews Dashboard	Storage · Pipelines · Tables + Vector (parallel)	One upload kicks two pipelines: sentiment analysis to a warehouse table + semantic embeddings to a vector collection. Same bucket, both pillars indexed.
Document Intake	Storage · Pipelines · Tables + Vector (parallel)	Similar fan-out shape: `document_triage` → Tables for routing logic, `text_embedding_index` → Vector for semantic recall. Every uploaded document feeds both decision-making and retrieval.

The pattern these recipes show

A K3 bucket isn’t just storage — it’s a knowledge instance. Each recipe demonstrates a different shape of multi-primitive composition:


   Storage bucket  ──upload──►  one or more ingest rules
                                          │
                                          ▼
                            ┌─────────────┴─────────────┐
                            │                           │
                            ▼                           ▼
                    Vector collection           Warehouse table
                  (semantic retrieval)          (structured rows, SQL)
                            │                           │
                            ▼                           ▼
                       Search RPC              Execute / Query RPCs

The fan-out happens because a bucket can have multiple ingest rules, and each rule targets one pipeline which writes to one destination. Want a Tables-bound copy of every uploaded doc AND a Vector-bound copy? Wire two pipelines on the same source, each with its own auto-generated rule.

Conventions used in these recipes

CLI-first where the CLI supports the action; API for everything else (rerank, multi-collection, multimodal, etc.).
Endpoint examples use staging https://k3.dev.dodil.io. Production: https://k3.dodil.io.
Bucket placeholder is kb-platform; vector engine + tables engine are both auto-on (vector engine on vector store create -m auto; tables engine auto-enabled at bucket-create time).
Consistent example collections / tables across recipes:
- docs — vector, pipeline-mode, text_embedding_index
- assets — vector, pipeline-mode, visual_embedding_index
- reviews_analyzed — table, pipeline-mode, review_analysis
- intake_triage — table, pipeline-mode, document_triage

Other useful pages

Per-primitive recipes — drill down into one primitive at a time:
Overview — the K3 platform map
Conventions — auth, headers, error envelope