Natural Language Processing

Semantic search, RAG retrieval, deduplication

Pick from MTEB retrieval and reranking evidence

Embeddings are infrastructure, not chat. Compare retrieval score, vector size, latency, license, language coverage, and serving cost before committing.

High-volume labels, moderation, routing, intent

Fine-tuned encoder or compact instruction model

For known label sets, small supervised models are easier to evaluate, cheaper to run, and less variable than a frontier chat model.

Machine translation for production

Dedicated MT system first; LLM second-pass only when needed

Route by language pair and domain. Use WMT, human adequacy/fluency review, terminology checks, and post-edit distance instead of a generic LLM pick.

Factual question answering over private data

RAG with citations, reranking, and answerability checks

The retrieval layer, citation discipline, and refusal policy usually decide quality. Benchmark with your documents and adversarial missing-answer questions.

Named entities and structured extraction

Fine-tuned token classifier, GLiNER-style model, or schema-constrained LLM

Use token classifiers for stable entity types and throughput. Use schema-constrained LLM extraction when entity definitions change often.

Summarization and rewriting

Current frontier LLM with factuality checks

Use a frontier generator when the output is prose. Still evaluate for dropped facts, invented claims, citation coverage, and controllable length.

Long-form reasoning or agentic workflows

Use the current frontier shortlist, then verify with task traces

Reasoning models change quickly. Do not hard-code an old model family; compare success rate, latency, tool-call reliability, and cost on your own workflows.