Knowledge (RAG) Settings

Most settings in this group point at something you build elsewhere. A Release does not hold a knowledge base; it references one. Build the vector first — see RAG Knowledge Base for creating one, choosing an embedding model, and feeding it by upload, crawl, or product-catalog sync.

Setting	Field	In the SDK?
Vectors / vector group	ragIndexes, ragVectorGroupId	✅
Recency weighting	ragRecency	✅
Max context chunks	rerankMaxChunks	✅
Retrieval gate	ragTrigger	✅
Citation display	ragContextDisplayMode	✅
Product sub-replies	productSubReplies	✅
Workflow pipelines	outputPipelineId, productOutputPipelineId	✅
Memory	memory	❌ — and see the caveat

Attaching knowledge

Two mutually exclusive ways to attach:

// Individual vectors
await divinci.releases.updateInWorkspace(workspaceId, releaseId, {
  ragIndexes: [{ id: "vec_abc" }, { id: "vec_def" }],
});

Or one vector group — a saved bundle with its own merge policy, set through ragVectorGroupId. Picking a group disables the individual list: it is one or the other, not both.

A group is created on the RAG vectors page (Groups tab) and carries:

Field	Default	Meaning
ragVectorIds	—	Ordered, minimum 2
mergeStrategy	interleave	interleave alternates chunks so every vector gets a voice; concatenate takes them in order
maxChunksPerVector	3	Caps how much any one vector contributes

Reach for interleave when the vectors are peers and you want balanced coverage; concatenate when the order expresses priority and you want the first vector to dominate.

Recency weighting

Lets newer content outrank older matches. Defaults, once enabled: weight 0.3, half-life 180 days — a document loses half its recency boost every six months. Unset at the Release level, resolution falls through to the vector’s own setting.

Worth turning on for changelogs, news, or policy docs where stale answers are wrong answers; leave it off for reference material where age is irrelevant and recency weighting would just distort ranking.

Max context chunks

rerankMaxChunks caps how many retrieved chunks survive the cross-index merge-rerank and reach the prompt.

Value	Effect
unset	Platform default of 8
n	At most n chunks
0	Disables the trim entirely

Lower means fewer prompt tokens and faster prefill, at the cost of fewer grounding sources. Note 0 does not mean “no chunks” — it means “no cap”, which is the opposite of what the number suggests.

Skipping retrieval

ragTrigger decides whether retrieval runs at all for a turn:

await divinci.releases.updateInWorkspace(workspaceId, releaseId, {
  ragTrigger: { mode: "heuristic", minChars: 4 },
});

heuristic skips retrieval for messages that plainly need no grounding — greetings, acknowledgements, chit-chat — saving both latency and prompt tokens. always (the default) always retrieves. minChars is the trimmed-length floor below which a message skips, and skipPatterns adds case-insensitive regexes on top of the built-in greeting set.

It never skips when grounding is required

If the Release enforces on-topic moderation (minimumContext or minimumTokens above zero), retrieval always runs regardless of ragTrigger — otherwise the gate would block every greeting for want of context it deliberately skipped fetching. See Safety → On-topic.

Citation display

Purely cosmetic — what the source citations under a reply look like:

ragContextDisplayMode	Shows
full	Title with hover preview
title	Title only
hidden	Nothing

This changes presentation only. It does not affect what was retrieved, what reached the model, or what it answered.

Product sub-replies

When a reply mentions something from your product catalog, the server can decorate the mention inline and spawn a follow-up product card. The catalog itself is built on its own page — imported from a file or synced from WooCommerce, Shopify, or Squarespace.

await divinci.releases.updateInWorkspace(workspaceId, releaseId, {
  productSubReplies: {
    enabled: true,
    maxProducts: 1,
    matchThreshold: 0.5,
    retroMatch: { enabled: true, mode: "name-mention" },
  },
});

retroMatch scans the model’s finished reply for products it named, even when the product’s chunk was not retrieved that turn. Its mode is the precision dial:

Mode	Behaviour
name-mention (default)	The product name must literally appear. Highest precision
bm25-full	Ranks over name + keywords + description. Higher recall, more false positives on common domain words
rag-candidates	Only products retrieval surfaced this turn. Never invents one the retriever did not see

Workflow pipelines

outputPipelineId and productOutputPipelineId point at visual node-canvas pipelines that post-process how retrieval and responses are assembled. Every Release ships with a simple RAG default; branch from it when you need logic the flat settings above cannot express.

Memory

This setting currently records intent and does nothing

release.memory is validated, saved, and carried forward when you fork a Release — but no generation path ever reads it. The live memory feature is ProfileMemoryService.compactThread(whitelabelId, …), which is keyed on the workspace, not the Release.

So setting a memory provider here changes no behaviour today. Configure memory at the workspace level (/white-label/{workspaceId}/memory), and treat the per-Release field as forward-looking configuration rather than something you can verify by testing a chat.

Gotchas

• A vector group and individual vectors are exclusive. Setting a group disables the list; do not expect the union.
• A group needs at least two vectors. One vector is just a vector.
• rerankMaxChunks: 0 disables the cap, not retrieval. If you want less context, use a small number, not zero.
• Citation display is not a privacy control. hidden stops rendering sources; it does not stop them being retrieved or sent to the model. To restrict what the model sees, change what is attached.
• Retrieval quality changes when de-identification is on. Queries run against the redacted message, so a corpus keyed on names, dates, or locations loses recall. Test retrieval with de-identification enabled, not before.

Related

• RAG Knowledge Base — creating and feeding vectors
• Caching & Performance — semantic caches over retrieval
• Safety — on-topic moderation, judged against retrieved context