Meridian Field Services: instructional example, not project evidence
One index at pilot: override-note search for the exception review (the US-3 surface: "find past overrides with similar reasons"). The label propagates: the vectors and the index are as sensitive as their sources.
| Field | Value |
|---|---|
| Source fields (cited) | Assignment.reason_code · Assignment.note[1] |
| Class · propagated | CONFIDENTIAL worker-conduct records (labels[2]); inversion reconstructs note text, so store access mirrors the source's: operations managers only (permission model[3]) |
| Chunking | One note, one vector; notes run 20 to 200 tokens; splitting buys nothing; rationale recorded |
| Metadata | assignment id · hub · reason_code · created_at (the filter fields) |
Model: the compact 384-dimension embedding model over the 1,536-dimension option. Quality on the held-out set differs by two points of Recall@10; storage and query cost by four times. Migration cost, stated: re-embedding the pilot corpus (60 thousand notes) is an afternoon and single-digit dollars, so the cheap model is reversible (embedding model[4]). Store: beside the relational data. The corpus sits far under the dedicated-store threshold and the filters are relational joins; the dedicated-store trigger is a ten-million-vector corpus or a hybrid-search need (vector store[4]).
Declared, not defaulted, and staged additions earn their place with triggers, so the future conversation has criteria. Anecdote is not a baseline.
| Config (exception-review search) | Value |
|---|---|
| Top-k · threshold | 20 · similarity 0.72 |
| Metadata filters · FIRST | hub · date range · reason_code: relational, cheap, and they cut the candidate set before any scoring |
| Hybrid (dense + lexical) | DEFERRED: reason codes are already a filter column, so the exact-identifier case is handled relationally; the trigger is free-text queries missing coded concepts past 10% of review sessions |
| Reranking | DEFERRED: first-stage precision meets the baseline; the trigger is Recall@10 below the gate on the refreshed set |
Baseline: Recall@10 0.91, MRR 0.78 on the golden query set (40 real review questions from M. Chen's team plus 60 synthetic, human-reviewed, register[5]); release gate: Recall@10 at 0.85. Retrieval-only, distinct from any generation eval, and none exists; the surface is retrieval-only at pilot (patterns[6]).
A stale note scores as high as a fresh one; similarity cannot tell. This section exists because of that.
| Rule | Value |
|---|---|
| Freshness · event-driven | A new override note embeds on write (the queue the worker already consumes, building blocks[7]); re-embed SLA: 5 minutes; the nightly re-index is the backstop only |
| Tombstones | The seven-year assignment archive transition (retention rules[2]) soft-deletes the vector (flagged, skipped at query time); asynchronous cleanup weekly; the retention rule drives the index, never the other way |
| Rot check | Nightly comparison of source write timestamps against vector timestamps; a gap past the SLA raises a ticket (alerting[8]) |
| Question | Owner | Answer by | Blocks |
|---|---|---|---|
| Do synthetic golden queries need union-visible review? (The corpus is worker-conduct text; shared thread with security open questions[9].) | N. Duval | Sep 30, 2026 | The golden set's refresh rule only |