Currently, online product research remains a time-consuming and difficult task. Differentiating trustworthy information remains challenging, and knowledge is siloed across disparate articles, reviews, Q&A threads, and product pages, requiring substantial work to analyze, organize, and apply this knowledge towards a shopping decision.
Generative AI and LLMs offer a clear opportunity to ease shopping research via a conversational interface. However, generalized, foundation LLMs lack specific knowledge about products, consumer sentiment, and detailed comparison criteria to provide actionable shopping advice that can drive purchases.
<aside> <img src="/icons/brightness-high_gray.svg" alt="/icons/brightness-high_gray.svg" width="40px" /> Product.ai is a knowledge assistant and shopping tool with deep understanding of shopping-related topics. Our mission with Product.ai is for it to feel more like talking to a real human expert with real, deep knowledge about you and the brands and products you’re considering.
</aside>
Demand.io’s core business relies on ShopGraph, our knowledge ontology about all things relating to e-commerce. ShopGraph powers our user-facing products and leads the industry with accurate, real-time information on millions of brands, sellers, products, UPCs, software, prices, offers, reviews, and decision criteria.
<aside> <img src="/icons/info-alternate_blue.svg" alt="/icons/info-alternate_blue.svg" width="40px" /> While no doubt many competitors will be building similar AI services to serve e-commerce customers, we believe our competitive advantage lies in the scale and quality of our knowledge ontology. We already have a substantial advantage over the industry in coupons data and knowledge, and we’re building a similar advantage in other areas of commerce including brand and product sentiment.
</aside>
Our model must handle with high degree of accuracy the following shopping use cases.
ShopGraph mission is to provide real-time, accurate knowledge and facts about all entities related to e-commerce. Data comes from crowdsourced inputs and web data ingested and interpreted by our ShopGraph LLMs. ShopGraph reaching scale is a requirement to deliver a reliable LLM user experience.
| Entity type | Subsystem | Description | Status | Size |
|---|---|---|---|---|
| Products | ProductsGraph | Product model, UPC | In production | 50k products, 500k UPCs (scaling up coverage now) |
| Categories | Taxonomy | Keyword, product type, category | In production | 2,000 nodes |
| Merchants | MerchantsGraph | Resellers, manufacturers, SaaS brands, single product brands | In production | 370,000 merchants |
| Offers | PromotionsGraph | Coupons, offers | In production | 20,000,000+ codes |
| Software | SoftwareGraph | SaaS, mobile apps, extensions, games, VR apps, AI services (models, plugins, agents) | Embedded in Products & Merchants, plan to separate | - |
| Blockchains | BlockchainsGraph | Tokens, cryptocurrencies, NFT projects, digital assets | Planned | - |
| Locations | PlacesGraph | Local businesses, local events, virtual locations, virtual events, AR places, AR events | Planned | - |
| People | PeopleGraph | Service providers, entrepreneurs, business owners, vendors, project leads, tastemakers, curators, creators. | Planned | - |
ShopGraph mission is to provide an accessible data store of all facts, specifications, and crucially, consumer sentiment about every entity related to e-commerce.
<aside> <img src="/icons/branch-create_blue.svg" alt="/icons/branch-create_blue.svg" width="40px" /> ShopGraph is structured as an ontology, comprised of interlinked, faceted taxonomies enabling classification, storage, and retrieval of categories, entities, and decision criteria of all topics relevant to e-commerce. We currently track all products, brands, manufacturers, resellers, and offers relevant to US consumers based on search volume.
</aside>
| Component | Description | Status |
|---|---|---|
| GPU server | Four Nvidia A100 GPUs for testing, fine-tuning, and deployment of internal LLMs. | Online, on premises |
| Open source, fine-tuned LLM | Evaluating Llama 2 (deployed) and Mistral models. | Deployed, in development |
| ShopGraph ontology | All taxonomies, metadata, relationships stored in RDBMS. Moving to graph database. | In production, in development |
| ShopGraph vector database | ShopGraph data in vector store (Pinecone or similar) | Planned |
| ShopGraph knowledge ingestion system | Set of workflows and LLMs capable of data ingestion, data scraping, crowsdsourced data combing, data analysis and interpretation, and organizing incoming information into ShopGraph ontology format. | In development |