Janady13/missionmatch / internal/modules/ai/memory.go

Source identity

Field	Recorded value
Source type	GitHub source
Original record	Janady13/missionmatch / internal/modules/ai/memory.go
Date	Not specified in source
Repository	Janady13/missionmatch
Revision	43d56387270ef4da4cabd794d97ca6a49909bc7d
Branch	w3-ai-staffing-recommendations
Path or archive member	internal/modules/ai/memory.go
Original SHA-256	bfe241262eb381116bfd4ae6b81a4cfa12cbc8694855a45198825b45d2fc5397
Git blob SHA	74ab31046bb150bcb0bc675dbb86cd4eed699394
Reading-copy SHA-256	bfe241262eb381116bfd4ae6b81a4cfa12cbc8694855a45198825b45d2fc5397

Original source ↗

Every recorded occurrence

These 3 records contain identical original source bytes. Their paths and revisions remain part of the archive.

Original record	Branch or archive	Revision / identifier	Original location
Janady13/missionmatch / internal/modules/ai/memory.go	w3-ai-staffing-recommendations	43d56387270ef4da4cabd794d97ca6a49909bc7d	Open original
Janady13/missionmatch / internal/modules/ai/memory.go	ai-intelligence-layer	5d5ddff70393a197c5f54a45280bb5dd42b24755	Open original
Janady13/missionmatch / internal/modules/ai/memory.go	w3-ai-staffing-recommendations	43d56387270ef4da4cabd794d97ca6a49909bc7d	Open original

Recovered source transcript

Download complete source text ↓124 lines · 4.0 KB

package ai

import (
	"context"
	"errors"
	"strconv"

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

	"github.com/missionworkforce/missionmatch/internal/platform/inference"
	"github.com/missionworkforce/missionmatch/internal/platform/tenant"
)

// Memory is the per tenant episodic and semantic memory, stored in the tenant's own database (ai_memory)
// and therefore physically isolated from every other tenant. Episodic memory records what happened (a
// craft note about a turn); semantic memory holds a recallable fact with an embedding. No demographic
// data may enter memory: every write is eeo guarded. Memory about a candidate carries the candidate id,
// so the right to erasure wipes it (ForgetCandidate), the same way anonymize wipes the resume vector.
type Memory struct {
	db     *tenant.DB
	client *inference.Client
}

// NewMemory builds the per tenant memory over the tenant database and the inference client (for semantic
// recall embeddings).
func NewMemory(db *tenant.DB, client *inference.Client) *Memory {
	return &Memory{db: db, client: client}
}

// Store writes one memory. candidateID may be nil for a non candidate craft note. content is eeo guarded.
// A semantic memory is embedded so it can be recalled by similarity.
func (m *Memory) Store(ctx context.Context, candidateID *uuid.UUID, kind, content string) error {
	if m.db == nil {
		return errors.New("ai: no database")
	}
	if kind != "episodic" && kind != "semantic" {
		return errors.New("ai: memory kind must be episodic or semantic")
	}
	if err := ForbidEEO(content); err != nil {
		return err
	}
	var vecLit any
	if kind == "semantic" && m.client != nil {
		emb, err := m.client.Embed(ctx, []string{content})
		if err != nil {
			return err
		}
		vecLit = vectorLiteral(emb.Embeddings[0])
	}
	return m.db.RunInTenantTx(ctx, func(ctx context.Context, tx pgx.Tx) error {
		_, err := tx.Exec(ctx,
			"INSERT INTO ai_memory (organization_id, candidate_id, kind, content, embedding) "+
				"VALUES (NULLIF(current_setting('app.current_org_id', true),'')::uuid, $1, $2, $3, $4::vector)",
			candidateID, kind, content, vecLit)
		return err
	})
}

// ForgetCandidate is the erasure hook: it deletes every memory about a candidate, so anonymize leaves no
// semantic trace in memory. It is called from the right to erasure path alongside the candidate
// anonymize. Idempotent.
func (m *Memory) ForgetCandidate(ctx context.Context, candidateID uuid.UUID) error {
	if m.db == nil {
		return errors.New("ai: no database")
	}
	return m.db.RunInTenantTx(ctx, func(ctx context.Context, tx pgx.Tx) error {
		_, err := tx.Exec(ctx, "DELETE FROM ai_memory WHERE candidate_id = $1", candidateID)
		return err
	})
}

// Recall returns the semantically closest memories to a query (RLS scoped, eeo guarded). It is the
// retrieval side of memory augmented generation.
func (m *Memory) Recall(ctx context.Context, query string, limit int) ([]string, error) {
	if m.db == nil || m.client == nil {
		return nil, errors.New("ai: memory recall needs a database and a client")
	}
	if err := ForbidEEO(query); err != nil {
		return nil, err
	}
	if limit <= 0 || limit > 50 {
		limit = 5
	}
	emb, err := m.client.Embed(ctx, []string{query})
	if err != nil {
		return nil, err
	}
	lit := vectorLiteral(emb.Embeddings[0])
	var out []string
	err = m.db.RunInTenantTx(ctx, func(ctx context.Context, tx pgx.Tx) error {
		rows, err := tx.Query(ctx,
			"SELECT content FROM ai_memory WHERE embedding IS NOT NULL ORDER BY embedding <=> $1::vector LIMIT $2",
			lit, limit)
		if err != nil {
			return err
		}
		defer rows.Close()
		for rows.Next() {
			var c string
			if err := rows.Scan(&c); err != nil {
				return err
			}
			out = append(out, c)
		}
		return rows.Err()
	})
	return out, err
}

// vectorLiteral formats a 384 dim vector as a pgvector text literal for the $::vector cast.
func vectorLiteral(v []float32) string {
	b := make([]byte, 0, len(v)*8+2)
	b = append(b, '[')
	for i, x := range v {
		if i > 0 {
			b = append(b, ',')
		}
		b = strconv.AppendFloat(b, float64(x), 'f', -1, 32)
	}
	b = append(b, ']')
	return string(b)
}