Language models and retrieval

A system that answers from your documents, with a number before every change.

Calling a modelMessages, temperature, tokens, cost, and structured output instead of free text.
Context engineeringWhat goes in the window and what comes out, and why longer is not better.
Evaluation firstFifty golden questions before any change, then a model as judge.
Vector searchWhat similarity brings together, where it misleads, and indexing in pgvector.
ChunkingBy paragraph, by heading, or overlapping, and what each does to retrieval.
Hybrid searchBM25 with vectors and a reranker, the right default in most cases.
Arabic retrievalNormalisation, diacritics, hamza forms and Arabic-Indic digits.
Retrieval metricsRecall, precision and MRR, then faithfulness to the source.
The system's limitsWhen the system declines, and scope as a query predicate not a prompt wish.
Project: an Arabic document assistantIt answers with a citation, and carries a number before and after every change.