The weekly strip · 63 issues · page 5 of 6
Building with AI
Hands-on patterns and code for people shipping AI features.
- Issue 38
10 October 2026 · Building with AI · 8 min read
Evals before agents: the smallest useful eval harness for a product team
An 85-line Python harness with a dataset, code checks, an LLM judge and a report, plus what to measure first and how much a pass rate on 100 cases can tell you.
- evals
- testing
- LLM
- Issue 37
10 October 2026 · Building with AI · 8 min read
Hybrid search: BM25 plus embeddings plus a reranker, and how to measure it
Why keyword search still earns a place, how reciprocal rank fusion merges it with embeddings, what rerankers cost, and an eval harness to run on your own data.
- hybrid search
- BM25
- reranking
- Issue 36
10 October 2026 · Building with AI · 6 min read
Milvus vs pgvector for a team of four
Milvus or pgvector? Index types, filtering behaviour, the 2,000-dimension index limit and the operational footprint, read from both projects' own docs.
- Milvus
- pgvector
- vector database
- Issue 35
10 October 2026 · Building with AI · 7 min read
Chunking is a product decision: sizes, overlap and the cost per query
Chunk size sets what the model reads, what you store and what each query costs. The evidence on size and overlap, a priced table, and a harness to test your own.
- RAG
- chunking
- embeddings
- Issue 34
10 October 2026 · Building with AI · 7 min read
RAG in 2026: when retrieval is optional, and when it is still the whole product
Every current Claude model has a 1M-token window and cached reads cost 5% of input. When to skip retrieval, when not to, with a priced 150k-token example.
- RAG
- long context
- retrieval
- Issue 30
10 October 2026 · Building with AI · 6 min read
1M context and 128k output: what the new Claude limits let you build
Every current Claude model has a 1M window and 128k output. The costs, the streaming rule for long generations, the 300k batch beta, and a worked bill.
- context window
- max output
- Claude
- streaming
- cost
- Issue 28
10 October 2026 · Building with AI · 7 min read
Cache read pricing compared: 0.025x to 0.1x across Claude, Gemini and GPT-6
Cache read multipliers, write fees, minimum prompt sizes and lifetimes for Claude, Gemini and GPT-6 in one table, plus break-even maths and a rule by traffic shape.
- prompt caching
- LLM pricing
- Claude
- Gemini
- GPT-6
- Issue 24
10 October 2026 · Building with AI · 7 min read
Confidence-gated routing: use model confidence as a second axis
Act, confirm or escalate: how to gate model answers on confidence, set thresholds by cost of error, and get a usable signal from Jev and from JSON-schema chat models.
- confidence
- routing
- human in the loop
- TypeSafe Jev
- structured output
- Issue 23
10 October 2026 · Building with AI · 8 min read
Noul: yes/no with a probability, and setting thresholds by the cost of error
TypeSafe's Noul returns the probability that a statement is true. How to phrase questions, pick thresholds from the cost of errors, and route the middle band to people.
- TypeSafe Jev
- Noul
- thresholds
- guardrails
- Python
- Issue 22
10 October 2026 · Building with AI · 7 min read
Jev's Score primitive: designing rubrics a model can actually use
How TypeSafe's Jev Score works: ordered levels, the score and confidence maths, why numbers in rubrics fail, and composite scoring to rank pull requests in code.
- TypeSafe Jev
- Score
- rubrics
- composite scoring
- Python
- Issue 21
10 October 2026 · Building with AI · 8 min read
Jev's Choice primitive: 255 options and the distribution you get back
How to write Choice criteria for TypeSafe's Jev, read the probability distribution, set confidence thresholds per question, and search deep taxonomies with a beam.
- TypeSafe Jev
- Choice
- classification
- confidence
- Python
- Issue 16
10 October 2026 · Building with AI · 6 min read
1M context by default: when long context beats retrieval, and when it doesn't
Most current Claude models include 1M tokens at standard pricing. A priced comparison of stuffing a 400k corpus versus retrieving 8k, and a decision table.
- long context
- RAG
- Claude
- prompt caching
- cost