Sift: Search That Answers From Your Own Content

Source codeem-ech/rag-siftops-knowledge-base Sift retail searchem-ech/rag-sift-retail-search Sift Drive, livesiftops-frontend-production.up.railway.app

Sift answers questions from an organization's own content and shows where each answer came from. The first version won a buildathon; two products followed: Sift Drive, a knowledge base over internal policy documents, and Sift's retail search, a shopping assistant that recommends only what a store sells.

RAGFastAPINext.jsQdrantOpenAIVector Search

policy documents indexed

departments covered

1st

of 30 teams at the buildathon

Top 5

of 30 at the venture bootcamp

Problem

Understanding the challenge and user needs

Keyword search returns documents that contain the words, and a general chatbot returns answers that sound right. Neither tells an employee what the travel policy says, or tells a shopper which shirt in this store suits a first date, with the source to prove it.

Approach

Strategy and methods used to address the problem

Sift Drive indexes 36 policy documents from five departments. Each document splits into passages, each passage becomes an embedding in a Qdrant vector database, and a question retrieves the closest passages before the language model writes an answer from them alone. Every answer lists its sources by file and page.

Sift Drive in use: a question about working from home returns the policy's requirements, each tied to a file and page, with the four source documents listed beneath.

Sift Drive search tab with the document tree at left — The sidebar lists every document by department; the Search tab returns the matching passages themselves.

Sift Drive assistant continuing a conversation with a follow-up question — The Assistant tab carries a conversation, and a question the documents cannot answer gets a plain statement of that instead of a guess.

What Sift Drive indexes

Department	Documents	Examples
Human resources	8	Remote work policy, employee handbook
Product and engineering	8	Architecture overview, deployment runbook
Finance	7	Expenses policy, travel policy
Security and IT	7	Password and MFA policy, incident response
Legal	6	Data retention, GDPR guide

Final Solution

The final design and outcome

The retail product applies the same retrieval to a product catalog. A retailer uploads a catalog, each store's data stays isolated from every other store's, and two script tags add search and a chat assistant to an existing storefront. The assistant reads the shopper's request, retrieves matching products, and answers only from that set.

The shopping assistant inside a demonstration storefront modeled on a fashion retailer's site. Asked for something to wear on a first date with a music lover, it recommends two band T-shirts from the catalog, with prices and links.

Product grid returned by Sift's semantic search — Semantic search returns products that fit the request, not only those whose names contain its words.

Code listing showing the two script tags that add Sift to a storefront — Integration takes a configuration object and one script tag on the store's page.

What runs where

Layer	Sift Drive	Sift retail search
Interface	Next.js 14	Next.js 15, embeddable widgets
API	FastAPI	FastAPI
Vector search	Qdrant	Qdrant Cloud, filtered per store
Records	PDF files	Supabase Postgres
Models	text-embedding-3-small, GPT-4o mini	text-embedding-3-small, GPT-4o

Impact & Validation

Results, learnings, and outcomes

The first prototype won the PwC, Lovable, and BETA DASH 3.0 buildathon among 30 teams and more than 100 participants. The retail product reached the final five of 30 teams at the IE Tech Venture Bootcamp.

Presentation

The deck, slide by slide

Title slide: SiftOps. A Retrieval-Augmented Internal Knowledge Base. Semantic search and grounded question answering over company policy documents. Group 2: Amaly Attia, Em Echeverria, Issam El Hoss, Marco De Palma, Nour Romdhane, Abdallah Lyagoubi. NLP Course Project, Option 1: Application Development.

Slide 1 of 10