RAG Retrieval — Published Products & Chat History

How the chat pipeline selects knowledge chunks for retrieval: published storefront products only, conversation-aware embedding, and link follow-up handling.

Related: RAG & chat API · Product sync · Conversation history · Vector call stack · Storefront product cards

---

Two history settings (env)

Env var	Default	Used for
CHAT_HISTORY_MESSAGES	50	LLM reply — how many prior user/assistant turns are sent to OpenRouter in buildPrompt()
CHAT_RAG_HISTORY_MESSAGES	8	RAG retrieval only — how many recent turns are prepended when embedding the search query

They are independent on purpose:

• 50 messages → better multi-turn answers (“what about that one?”, long threads).
• 8 messages → enough context for follow-ups like “what is the link?” without embedding an entire long thread every turn (lower cost, sharper vector search).

Both are set in .env / .env.sample.

---

RAG search query (conversation context)

Module: app/lib/rag-search-query.server.js → buildRagSearchQuery(message, historyForPrompt)

Before vector search, the app embeds text like:

Recent conversation:
Customer: what is the minimum price product you have?
Assistant: The minimum price product we have is Matching Separates, priced at $7,000.
Current customer message: and what is that product link?

Flow in app/lib/rag.server.js:

1. Load up to CHAT_HISTORY_MESSAGES from the thread (newest first, reversed to chronological).
2. ragSearchText = buildRagSearchQuery(message, historyForPrompt) — uses last CHAT_RAG_HISTORY_MESSAGES of that history.
3. getEmbeddingsBatched([ragSearchText]) → searchSimilarChunks(...).

The current user message is not duplicated in DB history when embedding (history is loaded before saving the current turn).

---

Link follow-up boost

When the customer asks for a link / URL / “that product” (isProductLinkFollowUp), and there is no focusProduct from a product card:

1. Parse product titles from the last assistant reply (extractProductTitlesFromAssistantHistory).
2. Resolve published catalog rows (findPublishedProductsByTitles).
3. Merge that product’s chunks ahead of vector results (mergeChunksProductFirst).
4. Inject Product from conversation context with Product page URL: … for the LLM.

Prompt rules (prompt-builder.server.js): when the customer explicitly asks for a link, include a markdown link from Store Context; do not fall back to the homepage only.

---

Published products only

Draft, archived, and unpublished products must not appear in RAG answers or product cards. Zero inventory is OK if the product is published.

Sync

Module: app/lib/product-storefront.server.js

• Shopify Admin search: status:active published_status:published (buildShopifyProductsSearchQuery).
• GraphQL field publishedAt stored on products.published_at (migration 20260523150000_add_product_published_at).

Ingestion

ingestProduct: if !isProductPublishedOnStorefront(product) → purge embeddings/chunks for that Shopify product ID and skip ingest.

purgeIneligibleProductsFromRag(shopId) runs at the start of each product sync to clean stale draft/unpublished embeddings.

Vector search

vector-search.server.js SQL excludes product-linked chunks unless:

lower(p.status) = 'active' AND p.published_at IS NOT NULL

Non-product chunks (FAQ, website knowledge, policies) are unchanged (product_id IS NULL).

Product cards

product-cards.server.js — catalog lookups use storefrontVisibleProductWhere() (active + publishedAt set).

Chunks

chunker.server.js — product overview chunks include Product page: {url} when available.

---

Deploy & re-sync

npx prisma migrate deploy   # includes products.published_at

Then run a full product sync in admin so:

• published_at is populated for live catalog items
• draft/unpublished products are purged from embeddings

---

Files

File	Role
app/lib/product-storefront.server.js	Eligibility, sync query, purge RAG
app/lib/rag-search-query.server.js	buildRagSearchQuery, link follow-up helpers
app/lib/rag.server.js	Orchestration: embed search text, conversation product boost
app/lib/vector-search.server.js	pgvector search with published-product filter
app/lib/ingestion.server.js	Gate + purge on ingest
app/lib/product-sync.server.js	Sync filter + publishedAt
app/lib/prompt-builder.server.js	Link-on-request + published-only rules