RAG-Powered Project Guidance Assistant for a Home-Improvement Retailer
Let's Connect
Overview
What we built
A 26-store home-improvement retailer had product know-how scattered across a PIM, a legacy CMS, thousands of PDFs and staff memory, so shoppers got a different answer in every store. We built a retrieval-grounded assistant that turns a plain-language project description into a step-by-step plan with sourced answers.
In plain terms: the retailer knew how to tile a bathroom, insulate an attic or wire a light fixture, that knowledge just lived in a dozen different places at once. Product documentation sat in a PIM, installation guides sat in a legacy CMS, supplier spec sheets filled thousands of PDFs, and a great deal of practical know-how existed only in staff memory. A shopper asking what they needed to tile a bathroom got a different answer depending on which of the 26 stores they walked into, and the website's search returned products, never guidance on how to use them.
We built a retrieval-augmented guidance assistant over a unified index of roughly 40,000 documents spanning product specs, installation guides and buying advice. A shopper describes a project in plain language and receives a step-by-step plan naming the exact products, quantities and tools required, with every answer citing its source document. Store associates use the same assistant on handheld devices in the aisles, so the answer a shopper gets no longer depends on who they happen to ask. Project-guidance sessions converted to purchase at 2.4x the rate of standard site search, and 93% of sampled answers cited a verifiable source in monthly accuracy audits.
The Problem
Project know-how scattered everywhere
The retailer's know-how was real, it was simply scattered. Product documentation lived in a PIM, installation guides lived in a legacy CMS, supplier spec sheets filled thousands of PDFs, and a large share of practical guidance existed only in the heads of experienced staff. No single system held the answer to a shopper's project question, so finding it meant knowing which of several places to look, and few people outside the business knew that.
The inconsistency showed up at the point of sale. A shopper asking what they needed to tile a bathroom could get a confident, correct answer from one associate and a vague or incomplete one from another, depending on the store and who happened to be on the floor. Across 26 stores, that meant the same question could produce 26 different possible experiences.
The website made the problem worse rather than better. Its search returned products, never guidance on how to use them, so a shopper who did not yet know which products they needed had nowhere to start. Project questions and product search were treated as the same problem, when a shopper planning a bathroom retile needed a plan before they needed a product list.
Know-how scattered across systems
Documentation lived across a PIM, a legacy CMS, thousands of PDFs and staff memory, with no single place holding the answer to a shopper's project question.
Answers varied by store
The same project question got a different answer in every one of the 26 stores, depending entirely on which associate happened to be asked.
Search returned products, not guidance
The website's search surfaced products but never explained how to use them, leaving shoppers who did not yet know what to buy without a starting point.
No step-by-step planning
There was no way for a shopper to turn a project idea into a concrete plan of products, quantities and tools before search even had a role to play.
What it was costing them
Every inconsistent answer sent a shopper away unsure whether they had what they needed, and every product-only search result left project questions unanswered. With know-how spread across a PIM, a legacy CMS, thousands of PDFs and staff memory, the retailer's own accumulated expertise was invisible to most of the shoppers walking into its 26 stores or browsing its site.
The Solution
Retrieval-grounded guidance assistant
We built a retrieval-augmented guidance assistant rather than trying to rewrite the retailer's documentation into one place. The assistant reads from a unified index of roughly 40,000 documents spanning product specs, installation guides and buying advice, so the underlying sources stay where they are while the assistant does the work of finding and combining the right ones for each question.
The output is a plan, not a link list. A shopper describes a project in plain language and receives a step-by-step plan with the exact products, quantities and tools required, and every answer cites its source documents so an associate or shopper can verify it. That citation discipline meant the assistant never presented a guess as settled advice.
We put the same assistant in the aisles, not just online. Store associates use it on handheld devices, so the guidance a shopper gets from a person is grounded in the same 40,000-document index as the guidance they would get typing into the website themselves. The answer to a project question stopped depending on which store or which associate a shopper happened to reach.
Key decisions
Index the existing sources, not rewrite them
The assistant retrieves from a unified index of roughly 40,000 documents across the PIM, CMS and spec sheets, so existing documentation stayed the source of truth.
Answer with a plan, not a product list
Responses give a step-by-step plan naming exact products, quantities and tools required, rather than leaving the shopper to translate guidance into a shopping list themselves.
Cite every source document
Every answer cites its source document, so associates and shoppers can verify guidance instead of taking a generated answer purely on trust.
Put the assistant in associates' hands
Store associates use the same assistant on handheld devices in the aisles, so in-store guidance draws on the same grounded answers as the website.
Keep it project-first, not search-first
Shoppers describe a project in plain language rather than searching for a product name, letting the assistant handle project questions the old product search never could.
Measurable Impact
What changed after launch
Shoppers who used the assistant bought at a noticeably higher rate. Project-guidance sessions converted to purchase at 2.4x the rate of standard site search, and the average basket on assistant-led orders came in 31% larger than the storewide online average, evidence that a full project plan sells more than a single product page does.
In-store, the assistant changed how associates worked. Associate product-question escalations to head office reduced by 57%, since associates could pull a grounded answer themselves rather than calling in. Trust in those answers held up under scrutiny too: 93% of sampled answers cited a verifiable source document in monthly accuracy audits.
Purchase conversion
Standard site search conversion baseline
Guidance sessions convert at 2.4x that rate
Basket size
Storewide online average basket
Assistant-led orders 31% larger
Head office escalations
Associates regularly escalated product questions
Escalations reduced by 57%
Answer reliability
No sourcing on associate or search answers
93% of sampled answers cite a verifiable source
Headline results
Project-guidance sessions converted to purchase at 2.4x the rate of standard site search
Average basket on assistant-led orders 31% larger than the storewide online average
Associate product-question escalations to head office reduced by 57%
93% of sampled answers cited a verifiable source document in monthly accuracy audits
Tech & Tools Used
What powered the build
Every tool below earned its place in this engagement. Here is the part each one played.
Python (FastAPI)
Serves the retrieval and answer-generation API that turns a shopper's plain-language project description into a sourced, step-by-step plan.
LangChain
Orchestrates the retrieval pipeline, pulling relevant passages from the document index and assembling them into the context the assistant answers from.
OpenAI API
Generates the step-by-step project plan and its product, quantity and tool recommendations, working only from the retrieved source documents for that query.
pgvector
Stores the vector embeddings for the roughly 40,000-document index, letting the assistant retrieve by meaning across product specs, guides and buying advice.
PostgreSQL
Holds document metadata and source citations alongside the vector index, so every generated answer can point back to the exact document it drew from.
Elasticsearch
Provides structured keyword search over the same document set, backing lookups where an associate or shopper searches for an exact product or spec term.
Next.js
Delivers the assistant's web interface, where shoppers describe a project and read the resulting plan alongside its cited source documents.
Redis
Caches frequently retrieved passages and in-progress project sessions, keeping responses quick for common questions across the 26 stores.
AWS S3 + ECS
S3 stores the underlying documents, spec sheets and PDFs while ECS runs the assistant's services, scaling with traffic across the site and the stores' handheld devices.
Ready to Build your Home Improvement Retail Business with RAG & Knowledge Systems
Ask Byte
Ask Byte
Typically replies instantly
just Now
Hi! I'm OrganByte's assistant. How can I help you today?
AI-generated content may be incorrect
