Techniques
Techniques
Use-case-driven data curation and cost-engineering techniques that make agents survive production. One file per technique. The library is open-ended: submissions welcome via the technique-submission issue template, he...
Required format (every entry)
- Problem: what breaks or costs too much without this technique
- Mechanism: how it works, concretely enough to rebuild
- Worked example: at least one production case (anonymized is fine); the scale it ran at
- Economics: token or cost math, before and after; order of magnitude acceptable
- When to use / when not to: decision criteria
- Failure modes and limits
- References: dated sources; maintainer knowledge cited from
knowledge/
Publication status
- Published: When a knowledge graph earns its keep, a contested-choice study comparing graph databases with embedding retrievers.
- Tested; results published as a research section, entry released from the publication hold on 2026-08-23: Multi-view (multi-card) embeddings with LLM-guided topic modelling. The pattern was tested in five rounds of experiments on public corpora (2026-08-19 to 2026-08-22); the results, including the negative ones, and the approach the evidence supports are published on the site's research pages (
/research, with the design at/research/recommended-approach). The technique entry, Multi-view embeddings with LLM-guided taxonomy, is published with a status note; it predates the experiments, and the research pages, not the entry, are the implementation guidance. - Planned, not yet guidance: Corrective RAG cascades; compaction and structured note-taking; distillation flywheels; semantic and prompt caching; deterministic routing tiers.
Source: techniques/README.md in the evidence repository behind this site.