Concept · living document

Search that speaks agent: LLM-enriched retrieval inside Cloze's architecture

Concept, Connor Simonson, July 2026. Tested on my own corpus before writing this up.

The problem

Agents don't remember data, they remember stories. The search box gets "that lender from the lake house closing" while the CRM holds field values, email signatures, and note fragments. Keyword search misses because the vocabulary doesn't overlap. The standard fix is semantic search with embeddings and a vector database, and that fix fights everything Cloze's architecture stands for: per-user isolated databases, local-first storage, no server-side pooling of customer data.

The method

Meta published a training-free alternative this spring (SIRA, arXiv 2605.06647, MIT-licensed). At ingestion, an LLM reads each record once and writes down the search phrases a user might type that the record itself never says. Those phrases get stored as plain text alongside the record. Retrieval stays ordinary keyword search. No embeddings, no vector index, no new infrastructure. On the standard retrieval benchmarks it beats the leading embedding model (0.691 vs 0.648 Recall@10).

I ported the method and ran it against my own 719-document knowledge base with a local model. Where queries were precise, it changed nothing. Where queries were vague and in plain language, the way a person actually asks, recall went from 0.825 to 0.950 and the right document moved toward the top of the list. The lift lives exactly where CRM search hurts.

Why this fits Cloze specifically

It's Maia-shaped. Cloze already runs AI over records at ingestion. Enrichment is one more pass, once per record, and the output is just text that syncs through existing pub/sub like any other field.
Per-user isolation survives. No shared vector index, no cross-customer model, nothing pooled. Each user's enrichment lives in their own database. The privacy story stays intact.
Query cost is zero. The LLM runs at ingestion. Search stays instant keyword matching, on device, offline included.

The audit angle

This is the part that matters in 2026. Post-NAR-settlement, "why did the system show me this" is no longer academic. Embedding similarity can't answer it; a number like 0.87 explains nothing to a compliance officer or a plaintiff's attorney. Enrichment phrases are readable text. You can inspect them, correct them, and lint them against fair-housing language lists before they ever enter the index. Explainable retrieval isn't a nice-to-have in real estate anymore, it's a defensible position.

Where it applies

Contact and history recall. "The inspector who found the radon issue in Carmel." The core Cloze search experience, made to match how agents actually remember people.
Closing document archive. Form Capture already extracts these documents. Enrichment makes years of transactions searchable: "the deal where the seller credited for the roof."
Listing search. Buyers say in-law suite; the MLS says accessory dwelling unit. Regional vocabulary (casita, granny flat) normalizes at ingestion. Works within conservative MLS infrastructure because it's still keyword search underneath. Enrichment output is auditable against HUD advertising-language guidance, which opaque semantic search can't offer.

What a test costs

A pilot is one enrichment prompt, one batch job over an existing index, and an A/B on search success rate. My 719-document run took under two hours on a laptop-class model. No migration, no new datastore, reversible by deleting a text field.

A living document: the source note is the single source of truth and this page re-renders from it on every site regeneration. Last built 2026-09-21.