Presets
A preset is a single named bundle of the settings, tools, prompts, and model choices that shape a CoreCube answer. Switching presets switches the whole runtime pipeline atomically — there is no half-applied state, no mix of "old retrieval, new prompt".
Presets exist because retrieval is multi-dimensional and coupled. Chunk size interacts with the embedding model. Reranker top-K interacts with the candidate pool. Hybrid weights interact with score floor. Tweaking one knob in isolation usually moves quality in unexpected directions. A preset captures a known-good combination that has been tested as a whole.
Admin Console → Configuration → Presets. Open a preset to edit its Pipeline, Ingestion, and Assignment tabs.
Seeded defaults and custom presets
Fresh installs include three editable seeded defaults:
| Preset | Purpose |
|---|---|
default-fast | Lean, low-latency answers with smaller retrieval pools. |
default-balanced | Fresh-install default. Adds query expansion and balanced retrieval pools. |
default-accurate | High-recall answers with wider retrieval and LLM chunk selection. |
You can edit these defaults directly, reset them to their shipped shape, or duplicate one into a custom preset before experimenting. Imported presets and presets created from scratch appear in the same preset library.
Anatomy of a preset
Every preset contains the same major areas, regardless of who created it:
| Area | Covers | Doc page |
|---|---|---|
| Pipeline | Query tools, default retrieval, chunk tools, answer prompts, answer model, action tools | Pipeline, tools, and prompts |
| Ingestion | Embedding model, chunking strategy, HNSW build params, ingestion-time prompt settings | Topic pages below |
| Assignment | Organization binding and optional compartment-specific overrides | This page |
The topic pages organize fields by what they do, not which tab they happen to live in:
- Pipeline, tools, and prompts — how a preset becomes a runtime pipeline
- Embedding — model selection and vector geometry
- Chunking — how documents become indexable units
- Retrieval — hybrid candidate generation, fusion, score floor
- Reranking — final-stage ranking, top-K, freshness
- HNSW vector index — pgvector ANN tuning
- System Prompts — prompt fragments and answer prompt composition
Each field's tooltip in the Admin Console links straight to its anchor on these pages — the Learn more → link inside the info popover takes you here.
How a preset takes effect
Presets are atomic on save, not on edit. While you are typing in the form, your changes live in a draft state — the running pipeline keeps using the previously-saved values. Pressing Save swaps the entire bundle in a single transaction.
The cost of switching depends on what you changed:
| Change category | Cost |
|---|---|
| Pipeline tools (query tools, default retrieval, chunk tools, action tools) | Instant after save/apply — the next query uses the composed pipeline snapshot. |
| Answer model and prompt fragments | Instant after save/apply — the next request uses the new answer model and composed prompt. |
| Retrieval-only knobs (top-K, weights, RRF k, score floor, freshness floor, reranker on/off) | Instant — applies to the next query. |
| Embedding model / dimensions | Re-embed of every chunk. Background job; old vectors continue to serve until the new ones land. |
| Chunking strategy / size / overlap | Existing chunks require a re-chunk transition; new ingests use the new shape. |
| HNSW M / efConstruction | Index rebuild. Background; current index continues to serve queries until the new one is ready. |
| HNSW efSearch | Instant — applied per query, no rebuild. |
CoreCube never silently re-embeds or re-chunks behind your back. Heavy operations are explicit background jobs you can monitor.
Drift
When the live runtime values diverge from what the active preset says they should be — typically because someone edited an underlying setting outside the preset flow, changed a shared prompt fragment, edited a shared tool, or upgraded into new preset-managed fields — the preset is in drift. The Admin Console shows a banner explaining what is out of sync. Re-saving or re-applying the preset reconciles the state.
Drift is informational, not destructive. The pipeline keeps running with whatever the live values are.
Related
- How retrieval works — the end-to-end flow these settings shape.
- Pipeline, tools, and prompts — how default presets, dynamic tools, and prompt fragments are composed.
- Embedding — start here when you want to change the model.
- Chunking — start here when documents seem to be split badly.
- Retrieval and Reranking — start here when results feel wrong.