story/internal/retrieval/memory.go

package retrieval

import (
	"context"
	"encoding/json"
	"fmt"

	"data-terminal-cloud/internal/developmentembedding"

	"github.com/jackc/pgx/v5"
)

type Memory struct {
	MemoryID        string          `json:"memory_id"`
	MemoryVersionID string          `json:"memory_version_id"`
	VersionNo       int64           `json:"version_no"`
	Content         string          `json:"-"`
	SourceRefs      json.RawMessage `json:"source_refs"`
	Sensitivity     string          `json:"sensitivity"`
	Relevance       float64         `json:"relevance"`
}

// SelectMemoryForContext fixes the tenant and subject boundary in the SQL
// itself. Callers cannot supply a free-form filter. The first production
// baseline admits only normal-sensitivity memories into a voice context.
func SelectMemoryForContext(ctx context.Context, tx pgx.Tx, tenantID, subjectID, query string, limit int) ([]Memory, error) {
	queryVector := developmentembedding.Literal(developmentembedding.Vector(query))
	return SelectMemoryForContextVector(ctx, tx, tenantID, subjectID, queryVector, limit)
}

// SelectMemoryForContextVector uses a vector generated before the caller's
// transaction. This keeps production embedding network calls out of the
// conversation creation transaction while retaining the SQL tenant, subject,
// consent and lifecycle boundaries below.
func SelectMemoryForContextVector(ctx context.Context, tx pgx.Tx, tenantID, subjectID, queryVector string, limit int) ([]Memory, error) {
	return selectMemory(ctx, tx, tenantID, subjectID, queryVector, limit, false)
}

// RESPOND_ONLY and closed items can answer an explicit question, but cannot
// seed a proactive greeting or follow-up prompt.
func SelectMemoryForUserQueryVector(ctx context.Context, tx pgx.Tx, tenantID, subjectID, queryVector string, limit int) ([]Memory, error) {
	return selectMemory(ctx, tx, tenantID, subjectID, queryVector, limit, true)
}
func selectMemory(ctx context.Context, tx pgx.Tx, tenantID, subjectID, queryVector string, limit int, responding bool) ([]Memory, error) {
	if limit < 1 || limit > 10 {
		return nil, fmt.Errorf("memory retrieval limit must be between 1 and 10")
	}
	if queryVector == "" {
		return nil, fmt.Errorf("memory retrieval vector is required")
	}
	rows, err := tx.Query(ctx, `
		SELECT cm.public_id::text,mv.public_id::text,mv.version_no,mv.content,mv.source_refs,
			cm.sensitivity,1-(mi.embedding <=> $3::vector) AS relevance
		FROM memory_index_entry mi
		JOIN curated_memory cm ON cm.id=mi.memory_id AND cm.tenant_id=mi.tenant_id AND cm.subject_id=mi.subject_id
		JOIN memory_version mv ON mv.id=mi.memory_version_id AND mv.memory_id=cm.id
		WHERE mi.tenant_id=$1::uuid AND mi.subject_id=$2::uuid
			AND mi.index_type='pgvector' AND mi.status='ACTIVE' AND mi.embedding IS NOT NULL
			AND cm.status='ACTIVE' AND cm.current_version_id=mv.id
			AND cm.proactive_policy<>'NEVER' AND ($5 OR cm.proactive_policy='ALLOWED')
			AND ($5 OR cm.memory_type<>'OPEN_ITEM' OR cm.item_status='OPEN')
			AND cm.sensitivity='normal'
			AND (cm.valid_from IS NULL OR cm.valid_from<=now())
			AND (cm.valid_to IS NULL OR cm.valid_to>now())
			AND EXISTS(SELECT 1 FROM consent_record cr
				WHERE cr.tenant_id=cm.tenant_id AND cr.subject_id=cm.subject_id
				AND cr.consent_type='personalized_memory' AND cr.withdrawn_at IS NULL
				AND (cr.expires_at IS NULL OR cr.expires_at>now()))
		ORDER BY mi.embedding <=> $3::vector,mv.confirmed_at DESC,mv.id
		LIMIT $4`, tenantID, subjectID, queryVector, limit, responding)
	if err != nil {
		return nil, fmt.Errorf("query eligible memory: %w", err)
	}
	defer rows.Close()
	items := make([]Memory, 0, limit)
	for rows.Next() {
		var item Memory
		if err := rows.Scan(&item.MemoryID, &item.MemoryVersionID, &item.VersionNo, &item.Content, &item.SourceRefs, &item.Sensitivity, &item.Relevance); err != nil {
			return nil, fmt.Errorf("scan eligible memory: %w", err)
		}
		items = append(items, item)
	}
	if err := rows.Err(); err != nil {
		return nil, fmt.Errorf("iterate eligible memory: %w", err)
	}
	return items, nil
}