Embedding Pattern Finder — Temporal Health Associations

What it does

Drop-in Go package that turns typed temporal health logs into frequency-framed behaviour→outcome associations — without causal or diagnostic claims. Includes 12h night bucketing (2am outcomes group with prior evening), probe-based affect classification, greedy embedding synonym clustering, minNightsWith noise gate, and 8 network-free tests. Wire to any Postgres session store + any Embedder implementation.

Why this beats rebuild

Embedding-driven backward-read pattern finder that buckets 2am outcomes with prior-evening actions, clusters synonymous behaviours, and frames co-occurrence without causal or diagnostic language.

Stack

Go, Postgres, pgvector, embeddings

Integration contract

Check the supported stack and delivery plan before you buy. Commands and full implementation details unlock with the bundle.

Preview

package patterns

// Finder is the query/pattern-finding side of the health graph.
// Pipeline:
//
//	load session nodes  →  split outcomes vs behaviours
//	classify each outcome-night negative/positive (embedding nearness)
//	cluster behaviour labels into neighbourhoods (embedding nearness)
//	per behaviour cluster, tally nights-with vs nights-bad  →  frequency framing
//
// Borrowed mechanism: embed text, compare by cosine. Fresh meaning: the
// action→outcome backward read, the night bucketing, and the safe framing.

import (
	"context"
	"fmt"
	"log"
	"math"
	"sort"
	"time"

	"github.com/google/uuid"
	"github.com/jackc/pgx/v5/pgxpool"
)

const (
	// nightOffset shifts an event back before bucketing to a date, so a 2am
	// "slept badly" outcome groups with the prior evening's actions rather