AI engineering

AI systems are software systems with an additional probabilistic component. Reliable behavior still depends on explicit contracts: which data enters the model, which vector space is used, how candidates are filtered, and how quality is measured.

Separate the pipeline

Retrieval quality should be decomposed into stages:

content → representation → index → candidate retrieval → filtering → ranking → response

Changing the model cannot fix a filter that removes the correct candidate. Increasing the retrieval limit cannot repair embeddings from incompatible vector spaces.

Articles

• Embedding search in practice — define vector compatibility, retrieval boundaries, filters, ranking, and evaluation before tuning similarity thresholds.

Planned series

• Chunking and document structure
• Hybrid lexical and vector retrieval
• Reranking and score calibration
• RAG evaluation datasets
• Agent tools, state, and failure recovery
• Tracing model calls without leaking sensitive data