Slim RAG Presets

From v3.5.0 Queria's RAG presets have been drastically simplified: they only carry sector identity (tags, domain terms, chunking) and no longer chat, retrieval, reranker, or LLM configuration. All pipeline logic has moved into DSL canvases.

Migration from legacy presets

If you come from Queria 3.1.x, your presets had 18+ sections (retrieval, reranker, llm, features, mmr, recency, aggregation, federated, etc.). These sections have been removed from the model. The equivalent settings now live in the Chat DSL canvases.

What presets contain today

interface SlimPresetConfig {
  chunking: {
    chunkSize: number         // 200-3000, default 2000
    overlapPercentage: number // 0-0.50, default 0.15
  }
}

Preset top-level fields remain:

• name, slug, description, icon, color
• sector (reference sector code)
• tags[]
• domainTerms[]
• isDefault, active

So today a preset serves two purposes:

1. Sector identity -- enabling the terms/tags and the sector integration during ingestion.
2. Chunking -- how to split documents into chunks (size and overlap).

Everything else lives in the DSL canvas of the pipeline (CHAT, INGESTION, SERVICE).

When to use which tool

I need to configure...	Where
How many documents to retrieve (topK)	retrieval node in Chat DSL
Which reranker to use	rerank node in Chat DSL
Which LLM, temperature, max tokens	llm_writer node in Chat DSL
Chunk size/overlap	RAG Preset > Chunking (this page)
Sector tags and domain terms	RAG Preset or Sectors YAML
Company-specific domain terms	Company sheet > Domain terms
Score threshold, MMR, diversity	retrieval node in Chat DSL
Citation policy	citation_pipeline node in Chat DSL
Federated search	external_tool node in Chat DSL

If unsure where a setting belongs, search the panel: the system always prefers canvas nodes over their "global" preset counterparts.

Sub-sector tab

The Preset page still has a sub-sectors tab where you can:

• Add additionalTags[] -- extensions of the sector's base tags.
• Add additionalDomainTerms[] -- fine vocabulary specific to the sub-sector.

Example: sector LEGAL, sub-sector tax-law:

• additionalTags: ["IVA", "IRPEF", "IRES"]
• additionalDomainTerms: ["agenzia entrate", "circolare", "interpello", "accertamento"]

These are merged with the sector's when a company with subSectorId = tax-law ingests or searches documents.

Create/edit a preset

Panel Sectors & Presets > RAG Presets > [new or edit]:

1. Identity: name, slug, description, icon, color.
2. Sector: choose the reference sector (LEGAL, FINANCE, ...).
3. Tags: domain keywords (e.g. for LEGAL: "contracts", "judgments", "regulations").
4. Domain terms: technical terms (e.g. "clause", "jurisdiction", "plaintiff").
5. Chunking:
   • chunkSize (200-3000): typical chunk size in characters. Default 2000.
   • overlapPercentage (0-0.50): overlap between consecutive chunks. Default 0.15 (15%).

Which chunk size to pick

• Discursive text (manuals, articles): 2000-2500 char, overlap 15-20%.
• Structured docs (contracts, codes): 1500-2000 char, overlap 10-15% -- article boundaries already help.
• Operational data (price lists, tables): not relevant, the operational chunker operates row by row.
• Short FAQs: 800-1200 char, overlap 5-10%.

Default preset and isDefault

Every sector has a preset marked isDefault: true that is automatically applied to new companies of the sector. To change the default:

1. Open the desired preset.
2. Enable isDefault.
3. The system removes isDefault from other presets in the same sector (unique-per-sector constraint).

When to rerun vectorization

Preset changes that require reingest of existing documents:

• Modifying chunkSize or overlapPercentage -> reingest needed to apply the new segmentation.
• Modifying domainTerms -> a sector-fields backfill is enough (no LLM, fast) to stamp the new terms on existing chunks.

Changes that do not require reingest:

• Name, description, icon, color.
• Add/remove tags.
• Change isDefault or active.

See Ingestion DSL > Reingest for operational procedures.

Preset deletion

A preset can only be deleted if no company and no topic references it. With active references, the system shows the list and blocks deletion.

Alternatively:

• Mark it active: false to hide it from new selections without breaking existing references.

Limits

• You cannot define a preset for multiple sectors at once: one preset = one sector.
• You cannot override chunking per-preset for a single document (only via custom INGESTION pipeline).
• There is no longer a CompanyPresetOverride concept: a company references the preset, does not override it. If a variation is needed, create a dedicated preset.

---

Queria v3.5.0 -- Slim presets (chunking + sector identity)