Recursive AI
onecontactDocs

The accurate tier

OneContact 1.0.0docs/models-accurate.md at 90204cd Markdown

All onecontact docs

From 1.1. The model packages described here ship in 1.1, after OneNLP 1.0. Version 1.0 runs on rules and markup alone; the seam in core that registers a model (ContactModels) is part of 1.0.

RecursiveAI.OneContact.Models.Accurate is the opt-in model package for harder text. It carries the same four heads as the fast tier on one frozen encoder: OneNLP's distilled multilingual student (RecursiveAI.OneNlp.Models.Multilingual.FrozenEncoder, six layers, width 384, 113 languages).

Using it

Reference the package; it registers itself through a source file the package adds to your project (no reflection, no call from you). AccurateModel.Instance is the loaded model; AccurateModel.Load(stream) loads a retrained head file.

  • Which package serves what. With several model packages referenced, the fast tier (.Models.Tier1, .Models) keeps normalizing single addresses and names, and the accurate tier serves the contact-span model that reads running text for ContactExtractor. With the accurate tier alone it serves addresses, names and spans. This is ContactModels.Register(..., normalizationRank: 0).
  • Rules first. Nothing changes about precedence: a rule hit is never overruled, and ParseDetail says whether a rule or the model decided each field.
  • No model referenced. An address the rules cannot decide fails with address.model-required, whose message names the packages to add, this one included.

Cost

The encoder is a transformer run on OneModel's CPU kernels, so the tier is slower and larger than the fast tier: about 140 MB of encoder files, about 0.8 GB resident once the encoder is loaded, and tens of milliseconds for a short text instead of fractions of one. Measured figures for maintainers are in docs/maintainers/models-accurate.md. Use the fast tier when you read single addresses and names in bulk.

OneNLP version

This package is built on OneNLP 0.4. A project that references it and a fast-tier package must resolve one OneNLP version; the fast tiers are retrained against 0.4 for that.

Retraining

recipes/accurate trains the heads with OneNLP's public Train over the frozen encoder, in resumable chunks on CPU: run.sh (address head), names.sh (name heads), spans.sh (span head). Its README lists the inputs. Provenance: src/OneContact.Models.Accurate/PROVENANCE.md.