New data extraction models #3

New issue

Open

opened 2026-07-23 02:53:35 +00:00 by benjaminsperry · 1 comment

benjaminsperry commented

2026-07-23 02:53:35 +00:00

Owner

@eliribble @ned

Per the discussion in #1, here are the two foundational documents for the modality-based knowledge graph refactor:

field-reference.md — Complete field spec for all 5 workflow domains (Feature Observation, Inspection, Treatment, Access Assessment, Safety Assessment). Each field lists its trigger word, description, and enum values pulled from the ontology. 24 fields total. Pure field reference for implementation.

utterances.yaml — 20 technician utterances mapped to structured fields, based on real field transcriptions from note-audio.csv:

Category	Count	Description
`structured`	10	Trigger-word dictation with full colon-delimited property:value pairs. Covers pools, gutters, yard drains, catch basins, fountains, depressions, ponds, containers, and dry features.
`natural`	10	Conversational speech matching real technician voice. Domain triggers only (`feature:`, `inspection:`, `treatment:`), no property-level trigger words.

Each utterance maps all 27 fields — absent values are excluded from the output.

How this connects to the modality proposal in #1:

The field reference defines what each mode needs to extract. The utterances provide training data for each mode's parser — structured colon-delimited pairs for the feature/treatment modes and natural conversational speech for the fallback.

Files attached:

20260722-field-reference.md
20260722-utterances-review.md

@eliribble @ned Per the discussion in #1, here are the two foundational documents for the modality-based knowledge graph refactor: **`field-reference.md`** — Complete field spec for all 5 workflow domains (Feature Observation, Inspection, Treatment, Access Assessment, Safety Assessment). Each field lists its trigger word, description, and enum values pulled from the ontology. 24 fields total. Pure field reference for implementation. **`utterances.yaml`** — 20 technician utterances mapped to structured fields, based on real field transcriptions from `note-audio.csv`: | Category | Count | Description | |---|---|---| | `structured` | 10 | Trigger-word dictation with full colon-delimited property:value pairs. Covers pools, gutters, yard drains, catch basins, fountains, depressions, ponds, containers, and dry features. | | `natural` | 10 | Conversational speech matching real technician voice. Domain triggers only (`feature:`, `inspection:`, `treatment:`), no property-level trigger words. | Each utterance maps all 27 fields — absent values are excluded from the output. **How this connects to the modality proposal in #1:** The field reference defines what each mode needs to extract. The utterances provide training data for each mode's parser — structured colon-delimited pairs for the feature/treatment modes and natural conversational speech for the fallback. Files attached: - `20260722-field-reference.md` - `20260722-utterances-review.md`

20260722-field-reference.md

9.3 KiB

20260722-utterances-review.md

24 KiB