Personalization in RAG

Three Layers of Personalization

Access Isolation: Per-User Namespaces

Mandatory when users have private data. Separate vector DB collections or namespaces per user.

from qdrant_client import QdrantClient
from qdrant_client.models import Distance, VectorParams, Filter, FieldCondition, MatchValue

client = QdrantClient(url="http://localhost:6333")

def ensure_user_space(user_id: str):
    coll = f"user_{user_id}"
    if not client.collection_exists(coll):
        client.create_collection(
            coll, vectors_config=VectorParams(size=1536, distance=Distance.COSINE)
        )

def retrieve_for_user(user_id: str, query_vec: list[float], k: int = 5):
    ensure_user_space(user_id)
    return client.search(collection_name=f"user_{user_id}", query_vector=query_vec, limit=k)

Alternative: single collection with hard metadata filter. Cheaper; requires a strict filter on every read:

user_filter = Filter(must=[FieldCondition(key="user_id", match=MatchValue(value=user_id))])
client.search(collection_name="kb", query_vector=query_vec, query_filter=user_filter, limit=5)

Namespace-per-user scales to ~10k users per cluster; filter-based scales to millions but every query must apply the filter or you leak data.

Preference Embedding

Embed the user profile once per session and blend with the query embedding.

import numpy as np
from openai import OpenAI

oai = OpenAI()

def user_profile_text(user_id: str) -> str:
    # Pull from your profile store
    role = profiles[user_id]["role"]            # "senior backend engineer"
    langs = profiles[user_id]["languages"]      # ["Go", "Python"]
    faves = profiles[user_id]["favorite_repos"][:5]
    return f"Role: {role}. Languages: {', '.join(langs)}. Interests: {', '.join(faves)}."

def blended_query(user_id: str, query: str, alpha: float = 0.8) -> list[float]:
    q_vec = np.array(oai.embeddings.create(model="text-embedding-3-small",
                                           input=query).data[0].embedding)
    u_vec = np.array(oai.embeddings.create(model="text-embedding-3-small",
                                           input=user_profile_text(user_id)).data[0].embedding)
    v = alpha * q_vec + (1 - alpha) * u_vec
    return (v / np.linalg.norm(v)).tolist()

Alpha: 0.8-0.9 is a sensible default. Query dominates; user bias nudges.

Personalized Reranking

Reranker sees the user context explicitly. More controllable than blended embeddings.

import cohere
co = cohere.Client()

def personalized_rerank(user_id: str, query: str, docs, top_n: int = 5):
    profile = user_profile_text(user_id)
    augmented_query = f"User: {profile}\nQuestion: {query}"
    res = co.rerank(
        query=augmented_query,
        documents=[d.page_content for d in docs],
        top_n=top_n,
        model="rerank-english-v3.0",
    )
    return [docs[r.index] for r in res.results]

For open models, use a cross-encoder with the same pattern.

Collaborative Filtering Signals

Borrow from recommender systems: users similar to this user engaged with these chunks.

# Precomputed: chunk_popularity[user_id] = {chunk_id: score}
# Similarity: cosine over normalized rating vectors.

def collab_boost(user_id: str, candidate_ids: list[str], k_neighbors: int = 20):
    neighbors = nearest_users(user_id, k=k_neighbors)
    boosts = {cid: 0.0 for cid in candidate_ids}
    for n, sim in neighbors:
        for cid, score in chunk_popularity[n].items():
            if cid in boosts:
                boosts[cid] += sim * score
    return boosts

def personalized_score(user_id, q, candidates, alpha=0.7):
    semantic = {d.id: d.score for d in candidates}
    collab = collab_boost(user_id, [d.id for d in candidates])
    return {cid: alpha * semantic[cid] + (1 - alpha) * collab.get(cid, 0)
            for cid in semantic}

Needs enough users + interactions to be worth it (> 1k active users, > 100k interactions).

Long-Term Memory: mem0 / Zep / Letta

mem0

from mem0 import Memory

mem = Memory()

mem.add("I prefer TypeScript over JavaScript and build with Vite.",
        user_id="alice", metadata={"kind": "preference"})

relevant = mem.search("What framework should I use?", user_id="alice", limit=5)

mem0 extracts structured facts, deduplicates, and stores in a graph + vector hybrid.

Zep (long-term memory + temporal reasoning)

from zep_cloud.client import Zep
z = Zep(api_key=...)

z.memory.add(session_id="alice_web", messages=[
    {"role": "user", "content": "I manage the platform team at Acme."}
])

mem = z.memory.get(session_id="alice_web")
# mem.relevant_facts includes entities, summaries, and timestamps.

Letta (MemGPT)

from letta import create_client, ChatMemory
client = create_client()
agent = client.create_agent(
    name="alice_agent",
    memory=ChatMemory(human="Alice, senior backend engineer, prefers Go.",
                      persona="You are a helpful technical assistant."),
)

Letta manages a hierarchical memory (core, archival, recall) with an internal memory-management LLM loop.

Graph-Based User-Entity Memory

Facts as (subject, predicate, object, timestamp). Retrieval by entity and relationship.

# Neo4j via Cypher
"""
MERGE (u:User {id: $user_id})
MERGE (e:Entity {name: $entity})
MERGE (u)-[r:RELATES {kind: $rel}]->(e)
SET r.first_seen = coalesce(r.first_seen, datetime()),
    r.last_seen = datetime(),
    r.count = coalesce(r.count, 0) + 1
"""

def expand_query_with_entities(user_id: str, query: str):
    """Expand the query with the user's most-connected entities from the graph."""
    res = neo4j.run("""
        MATCH (u:User {id: $u})-[r:RELATES]->(e:Entity)
        RETURN e.name AS name, r.count AS c
        ORDER BY c DESC LIMIT 10
    """, u=user_id).data()
    entities = [row["name"] for row in res]
    return f"{query} (user interests: {', '.join(entities)})"

Privacy & Compliance

Encryption at rest

Metadata that contains PII must be encrypted. Vector DBs do not typically provide per-field crypto; use an application-layer wrapper.

from cryptography.fernet import Fernet

cipher = Fernet(KEY)

def store(user_id: str, vec, meta: dict):
    safe_meta = {
        "user_id": user_id,
        "ciphertext": cipher.encrypt(json.dumps(meta).encode()).decode(),
    }
    client.upsert(...)

def retrieve(user_id, q_vec):
    res = client.search(...)
    for p in res:
        p.payload["meta"] = json.loads(cipher.decrypt(p.payload["ciphertext"]))
    return res

GDPR Right-to-Erasure

Namespace-per-user makes deletion one operation:

def delete_user(user_id: str):
    client.delete_collection(f"user_{user_id}")
    profiles.pop(user_id, None)
    mem.delete_all(user_id=user_id)
    neo4j.run("MATCH (u:User {id:$u}) DETACH DELETE u", u=user_id)

With filter-based isolation, you also need:

client.delete(
    collection_name="kb",
    points_selector=Filter(must=[FieldCondition(key="user_id", match=MatchValue(value=user_id))]),
)

Audit: record the deletion event (ID, timestamp, triggering request) to an immutable log for regulatory proof.

Differential Privacy for Collaborative Features

If collaborative signals leak user behavior back to other users, add calibrated noise.

import numpy as np

def dp_boost(boosts: dict, epsilon: float = 1.0):
    sensitivity = 1.0  # cap per-user contribution
    scale = sensitivity / epsilon
    return {k: v + np.random.laplace(0, scale) for k, v in boosts.items()}

Epsilon 1-5 is typical for recommender systems; stricter in regulated sectors.

Data Minimization

Do not store raw queries indefinitely. Extract facts, discard raw text after N days.

# Daily job:
# 1. Extract facts from query_log via LLM.
# 2. Store facts in user profile.
# 3. Delete raw query log entries > 30 days old.

Evaluation of Personalization

• A/B test personalized vs non-personalized retriever.
• Per-user lift: average top-K relevance delta per user.
• Cold start: how does the system behave with no profile data?
• Fairness: does personalization disadvantage a protected group?

Anti-Patterns

Production Checklist

• Namespace-per-user or hard filter on every retrieval
• Right-to-erasure endpoint deletes across vector DB, profile, memory store, graph
• Deletion events logged immutably with timestamp + request ID
• PII metadata encrypted at rest
• Per-user token / query rate limits
• User profile store (Postgres) with versioning
• Long-term memory store (mem0/Zep/Letta) evaluated
• Collaborative signals gated behind minimum activity threshold
• Differential privacy on cross-user aggregates if exposed
• Cold-start experience designed (default profile, onboarding questions)
• Personalization A/B tested vs baseline
• Fairness audit across protected groups
• Raw query retention policy (delete after N days, keep extracted facts)