Patterns over prompts - feeding AI existing curriculum content as examples dramatically outperformed adding more constraints to prompt instructions; generate-many-filter-to-best approach (create multiple episodes, AI evaluators select highest quality) more effective than trying to generate perfect content on first try; curriculum foundation ensures level-appropriate, grammatically sound content

Enriching table metadata with business terminology, few-shot examples from domain experts, and multi-stage retrieval algorithms transforms generic GPT-4 into domain-expert Text-to-SQL capable of production use - simply using GPT-4 alone produces queries lacking company context, ignoring data policies, and suffering hallucinations, but combination of (1) augmented documentation with detailed column descriptions and business glossaries, (2) sophisticated search algorithms refining questions and selecting relevant examples, and (3) ReAct prompting reasoning step-by-step while dynamically retrieving context enables query quality that employees trust for actual work; 'garbage in, garbage out' principle applies where foundation model performance capped by input quality

Knowledge distillation dramatically reduced costs while maintaining quality - teacher model (GPT-4o) trained smaller student (GPT-4o-mini) to achieve same quality with much shorter prompts

Agent abstraction decoupling input/output from implementation allows multiple variations honoring same contract but differing in complexity/models/optimization - maintaining strong inter-agent compatibility and reusability for rapid development; if solution works in developer's local environment for single instance, platform automatically scales to eBay's industrial needs serving hundreds of millions across billions of listings; NRT distributed queue-based messaging smooths user activity peaks/valleys for consistent controllable throughput and better GPU utilization

Documentation is foundation for LLM discovery - without high-quality docs (increased from 20%→90% coverage), LLM chatbot cannot work effectively; four distinct search categories (exact, partial, inexact, semantic) require different solutions - keyword search handles 75%, LLM handles semantic; incremental approach (Elasticsearch→Docs→LLM) validated each step before next; leveraging existing Glean tool accelerated go-to-market vs building from scratch

Single-pass generation essential to avoid accuracy degradation - 90% accuracy compounds to 59% over 5 chains, so chaining multiple LLM calls fails; few-shot prompting outperformed zero-shot and chain-of-thought; accepting 'good enough' outputs (flexible interpretation like user typing 'slow') serves users better than rigid correctness; LLMs are engines for features not standalone products; schema filtering by 7-day activity handles context limits for 5,000+ field customers

Automating quality evaluation with LLMs enables continuous, rapid iteration on search improvements that would be impossible with manual review - by reducing evaluation time from days/weeks to hours, LinkedIn can experiment with search enhancements and measure impact quickly, dramatically accelerating search quality improvements; slow feedback loops prevent experimentation and innovation, but GPT-powered evaluation provides consistent, fast assessment enabling continuous improvement at platform scale serving billions

Table documentation quality trumps model sophistication - weighting embeddings toward table metadata increased search hit rate from 40% to 90%, proving data governance is the primary bottleneck for Text-to-SQL performance; real-world deployment revealed table discovery (finding right tables among hundreds of thousands) and metadata quality (accurate descriptions of table purpose and column meanings) far more critical than prompt engineering or model selection; benchmarks like Spider fail to capture this as they treat small number of pre-specified well-normalized tables as given