Thoughts, articles, and resources on software development, SaaS, and digital growth.
Every technique for reducing LLM hallucination sounds compelling in a blog post. The harder question is which ones are actually worth building for your specific system, your specific failure modes, and your specific cost constraints. Here is how we worked through that for Forseti.
The debate between RAG and long context is mostly theoretical. Here is what building a production system that uses both deliberately actually taught us about when each approach earns its place.
Changing chunk size fixes one retrieval and breaks another. Setting temperature to zero does not guarantee determinism. The behaviour is unstable, context-dependent, and often surprising. Maintaining a generative RAG system is not pure engineering — it is a craft that requires both technical currency and deep subject matter expertise.
Embedding a user's question and searching for similar text sounds right. For formal documents like regulations, contracts, or technical standards, it systematically retrieves the wrong content. Here is the problem, and a technique that solves it.
Self-healing pipelines fix data that arrived broken. But what about data you never knew you needed? Here is how to build a system that detects knowledge gaps at query time and ingests the missing source before returning an answer.