Node.js Semantic Search for Product Support — Stable Chat Contracts for Messy Catalogs

Description and use

An in‑app chatbot uses embeddings for semantic search over messy product descriptions, applies tenant filtering before the chat model call, and tracks usage per stage, providing a single credential and bill.

Problem

In‑app customer‑support chatbots need to retrieve relevant catalog items from messy product descriptions while keeping per‑tenant cost visibility and a stable API contract across provider changes.

Outcome

The chatbot now handles paraphrased queries with semantic search, enforces tenant boundaries before prompting the model, and tracks cost per tenant across ingestion, query, retrieval, and generation. A single credential and bill simplify operations while allowing provider flexibility.

Workflow

• Normalize each catalog description and attach tenant ID
• Create and store an embedding for each description
• Embed incoming support question
• Retrieve top candidate records within tenant boundary
• Send selected records and question to chat model for answer

Evidence

Embeddings give the description and the incoming question comparable semantic representations.

Source

Similarity is not authorization. The tenant filter must run before the result reaches the prompt.

Source

Use one ledger row per stage. An import event needs an import ID. A live question needs a request ID.

Source

The single-key, one-bill model covers 295 routes across 20 modules.

Source