TECH Signal 649 2 feeds carried it
LLMs reportedly hallucinate fake classifications to map queries to real product taxonomies
↗
A method uses small language models to generate plausible but fake product categories, then matches them to real ones via embeddings.
This approach avoids sending large taxonomies to LLMs, reducing cost and latency. It shifts classification from strict schema enforcement to approximate matching, which may trade precision for scalability. Engineers building search or recommendation systems can adopt it without retraining models.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Generates fake but plausible product categories instead of constraining LLM outputs to a fixed schema.
Uses embeddings to map hallucinated categories to real ones, reducing input size and model complexity.
Works with smaller, cheaper models and avoids shipping large taxonomies with each query.
THE READ
What the cluster adds up to.
The method inverts the usual classification workflow. Instead of forcing an LLM to pick from a predefined list of categories, it asks the model to invent new ones that fit the query. This avoids the need to send the full taxonomy with every request, which can be impractical for large datasets like Wayfair’s 500+ categories. The trade-off is that the LLM’s output may not align perfectly with the real schema, requiring a second step to resolve it.
Embeddings bridge the gap between hallucinated and real categories. By computing vector representations of both, the system can use similarity search to find the closest match in the real taxonomy. This step is computationally cheap and can run in-memory, making it suitable for high-throughput applications. The approach relies on the assumption that the LLM’s invented categories will be semantically close to the real ones, which may not hold for edge cases or ambiguous queries.
Cost and scalability drive this design. Smaller models like MiniLM are sufficient for generating embeddings and hallucinating categories, reducing inference costs compared to larger models. The method also avoids the overhead of structured outputs, which can be slow or error-prone when dealing with hundreds of categories. However, it introduces a dependency on the quality of embeddings, poor matches could lead to misclassification, and the system may struggle with taxonomies that lack clear semantic structure.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER