Rather than indiscriminately feeding unstructured data into LLMs, invest in transforming it into structured formats to avoid hallucinations and improve reliability.
Define and implement an ETL pipeline by extracting raw use-case data, transforming it into structured themes, deduplicating and merging, and loading it into an analysis workflow.
Build mobile apps that integrate on-device LLMs with vectorized schema metadata to guide users through configuration tasks without server dependencies.
They initially built a langraph with supervisor and sub-agents triggering JSON schemas, but prompt engineering complexity and reliance on frontier models proved impractical.
Extensive AI agent architectures and heavy prompt engineering can be outperformed by a simple prompt plus vectorized retrieval, reminding engineers not to over-engineer.
Use a lightweight on-device LLM paired with a retrieval augmented generation pipeline over vectorized schema metadata to handle interactive user inputs efficiently.
Transform JSON schema fields with metadata into vector embeddings and use a simple LLM to retrieve and fill the right fields, bypassing complex prompt engineering.
Use prompt engineering over streaming transcripts to map free-form conversational descriptions—like “windows look a bit shabby”—to specific, metadata-annotated schema fields in a JSON inspection template.
Leverage a voice assistant to automate property condition assessments by converting spoken observations directly into industry-standard inspection reports.
Implement a chat-based LLM pipeline that processes each turn of streamed natural-language input, identifies relevant fields in a predefined JSON schema, and fills an inspection report in real time.
While building a real estate inspection app, the team pivoted from using complex AI reasoning to straightforward vector search on-device, simplifying the implementation.
Use lightweight vectorization of JSON schemas and data dictionaries to map natural language inputs to structured outputs via vector search instead of complex reasoning.
LLM search driven by semantic similarity in embedding space is far more sophisticated and efficient than basic keyword matching, enabling nuanced query understanding.
Use UMAP to reduce high-dimensional embeddings (e.g., 784-dim FashionMNIST) into 2D/3D to visualize and identify semantic clusters such as trousers, dresses, and footwear.
