How Allspice Improved Ingredient Matching from 20% to 97% with Pinecone

Allspice is a food technology company building a comprehensive kitchen operating system that serves both consumers and recipe publishers. On the consumer side, the platform helps home cooks discover recipes, manage pantry inventory, and generate automated shopping lists. For publishers, Allspice provides interactive tools that increase engagement and unlock revenue streams beyond traditional display advertising. At the center of both experiences is the ability to understand food the way a human does — recognizing that “one bunch of cilantro” and “fresh cilantro, chopped” are the same ingredient despite entirely different wording.

As Allspice expanded recipe importing into a primary product feature, it hit a fundamental wall. Ingredient data is inherently messy: modifier-heavy phrases, spelling variations, parsing inconsistencies, and informal descriptions make deterministic matching unreliable. The team had built on a NoSQL stack using Google Cloud Firestore and Typesense, which worked well for traditional structured search but could not bridge the semantic gap between how ingredients appear in source recipes and how they are stored in Allspice’s proprietary database. Attempting to use Typesense’s vector capabilities created additional problems — storing large embeddings alongside small documents degraded overall system performance. With ingredient matching sitting at roughly 20% accuracy, recipe importing was not viable as a product feature.

Allspice turned to Pinecone as a dedicated, purpose-built vector database. The team embedded their proprietary ingredient database of approximately 10,000 entries using OpenAI’s text-embedding-3-large model and immediately validated the approach. Pinecone’s serverless, managed infrastructure kept semantic search fully decoupled from the rest of the stack, allowing vector workloads to scale independently without rearchitecting existing layers. What began as a single ingredient-matching pipeline expanded iteratively: recipe-level embeddings enabled a “more recipes like this” feature, fuzzy recipe search added flexible retrieval alongside structured filtering, and Pinecone-hosted Llama embeddings supported chatbot data normalization by mapping free-form user inputs to structured internal representations. The platform now manages approximately 110,000 total recipe and ingredient embeddings.

The results transformed the product. Ingredient matching accuracy jumped from 20% to 97%, moving the recipe importing pipeline from unusable to production-ready. With that foundation in place, Allspice was able to launch one of its core features — a capability that had been blocked entirely before the vector layer existed. The team set up Pinecone and validated a working pipeline in a single afternoon, a speed that was essential for a startup where idea validation determines what ships. User satisfaction with recipe search improved significantly, and publishers gained new engagement surfaces that increased time on site and enabled revenue from grocery exports, subscriptions, and affiliate commissions.

Allspice is now expanding Pinecone’s role into its AI and conversational systems. Planned work includes FAQ classification and retrieval to match user questions against publisher-approved content, and broader use of vector retrieval to reduce LLM token usage as chatbot query volume grows. The stack — Pinecone for semantic retrieval, Gemini 2.5 and GPT-4.1-mini for generation, and Firestore for document storage — keeps Allspice model-agnostic while grounding its AI features in reliable, high-accuracy retrieval.