Designed for Claude Code, also compatible with Codex
LangChain Enterprise RBAC (Python)
Overview
A B2B SaaS team shipped their first RAG feature for two tenants. The factory
code looked innocent: build PineconeVectorStore once at module import with
namespace="acme-corp" (the first tenant), convert it to a retriever, store
it in a module global, reuse on every request. Six weeks later tenant "Initech"
went live. Their first search returned three documents from Acme Corp.
The singleton retriever had captured the Acme namespace at process start.
RunnableConfig.configurable["tenant_id"] was being passed in — but the
retriever never read it, because the filter was baked in. Every request for
every tenant hit the same Pinecone namespace. Security review caught it three
days later and put a hold on the SOC2 renewal. This is pain-catalog entry
P33, the single most common cause of cross-tenant leak in LangChain 1.0
production.
This skill fixes it with four workstreams:
P33 — retriever-per-request factory — build the retriever inside the
chain or agent invocation, keyed by tenant_id from RunnableConfig. Never
at module scope. Unit-test with two tenants and assert non-overlap.
Role-scoped tool allowlist — build the agent per-request with only the
tools the current user's role permits. Forbidden tools are not passed to
create_agent at all, so the model cannot call them even if it tries.
Per-tenant rate limit + budget — scope the InMemoryRateLimiter (or a
Redis-backed equivalent) by tenant_id, and check a per-tenant USD budget
before invoking the model.
Structured audit log — JSON log with user_id, tenant_id,
chain_name, tools_called, cost_usd, outcome, emitted in both success
and failure paths. Ships to SIEM or BigQuery.
Two failure patterns anchor this skill: import-time retriever binding (P33)
and missing audit log on tool failure (the try block logs on success but
the except branch re-raises without emitting, so incident response has no
record of denied tool calls). Pinned: langchain-core 1.0.x,
langgraph 1.0.x, langchain-anthropic 1.0.x, langchain-openai 1.0.x,
langchain-postgres 0.0.15+ (for PGVector RLS), ,
. Pain-catalog anchors: , P18, P24, P31, P37.
At least one vector-store backend: pinecone-client, langchain-postgres
(PGVector), chromadb, or faiss-cpu (single-tenant only — see §Step 2)
A structured logging sink: stdout JSON for local, Cloud Logging / Datadog /
Splunk HEC / BigQuery streaming insert for production
A tenant authorization claim in every request (JWT tid claim, session
cookie, or header — the auth boundary is out of scope for this skill but
assumed correct)
Instructions
Step 1 — Build the retriever per-request, never at import
Move retriever construction inside the chain or agent invocation, keyed by
config["configurable"]["tenant_id"].
from langchain_core.runnables import RunnableConfig, RunnableLambda
from langchain_pinecone import PineconeVectorStore
# WRONG — retriever bound at import time with first tenant's namespace.# RETRIEVER = PineconeVectorStore(index_name="rag", namespace="acme-corp",# embedding=emb).as_retriever(search_kwargs={"k": 4})# RIGHT — factory called per-request, reads tenant from RunnableConfig.defretriever_for(config: RunnableConfig):
tenant_id = config["configurable"]["tenant_id"] # required, no defaultifnot tenant_id:
raise PermissionError("tenant_id missing from RunnableConfig")
store = PineconeVectorStore(
index_name="rag",
namespace=tenant_id, # P33 fix — namespace per-invocation
embedding=emb,
)
return store.as_retriever(search_kwargs={"k": 4})
defretrieve(inputs: dict, config: RunnableConfig):
return retriever_for(config).invoke(inputs["query"])
chain = RunnableLambda(retrieve) | prompt | model
result = chain.invoke({"query": "..."},
config={"configurable": {"tenant_id": "initech"}})
No default on tenant_id — a missing tenant must be a hard error, not a silent
fallback. See Retriever-per-request for
factory lifecycle and PGVector RLS / Chroma / FAISS adapters.
Step 2 — Pick a vector-store isolation primitive
Store
Isolation primitive
Per-tenant latency
Max tenants
Safety notes
Pinecone
namespace=tenant_id per query
~40ms p50 (shared index)
100,000+ per index
Namespace is the documented isolation boundary; still apply metadata {"tenant_id": tid} as defense in depth
PGVector
Postgres row-level security (RLS) on tenant_id column
~20ms p50 (HNSW index)
Bounded by Postgres row count
Use SET LOCAL app.tenant_id = :tid per transaction; RLS policy USING (tenant_id = current_setting('app.tenant_id'))
Good isolation at small scale; collection creation is synchronous — provision lazily but cache the handle per-request
FAISS
In-process index-per-tenant
~5ms p50 (in-memory)
10-50 practical limit
Poor fit for multi-tenant SaaS — cannot shard across processes, reloads on every deploy, no durable filter. Use for single-tenant evaluation only
Pinecone scales highest. PGVector with RLS is the strongest primitive when
isolation must be auditable at the database layer (RLS is enforced by the
server even if application code is bypassed). Chroma is fine for ≤1,000
tenants. FAISS is not a multi-tenant production choice — document so future
engineers do not adopt it. See
Vector-store isolation for the PGVector
RLS DDL and Chroma lifecycle.
Step 3 — Build the agent per-request with a role-scoped tool allowlist
Never bind every tool to every agent and trust the model to pick the right one.
Build the agent inside the request handler with only the tools the user's role
permits.
from langgraph.prebuilt import create_agent
# Map role -> allowed tool names. Owned by IAM config, not the skill.
ROLE_TOOLS: dict[str, set[str]] = {
"viewer": {"search_docs"},
"editor": {"search_docs", "create_note"},
"admin": {"search_docs", "create_note", "delete_note", "export_audit"},
}
ALL_TOOLS = {t.name: t for t in [search_docs, create_note, delete_note, export_audit]}
defagent_for(user_role: str, tenant_id: str):
allowed = ROLE_TOOLS.get(user_role, set())
tools = [ALL_TOOLS[n] for n in allowed if n in ALL_TOOLS]
# Forbidden tools are not passed in — the model never sees them.return create_agent(model, tools=tools)
agent = agent_for(user_role="viewer", tenant_id="initech")
# viewer has no delete_note -> the agent cannot call it.
Add a denylist for dangerous argument patterns (SQL with DROP / TRUNCATE,
shell with rm -rf / sudo, URLs to internal metadata endpoints) via a
pre_model_hook or tool wrapper — allowlist bounds which tools run,
denylist bounds what arguments they accept. See
Role-scoped tool allowlist.
Step 4 — Scope rate limits and budgets by tenant
Per-tenant limits prevent one tenant's runaway job from exhausting shared model
capacity. Rate-limit detail is covered in langchain-rate-limits; cost-budget
detail in langchain-cost-tuning. The only RBAC-specific requirement here:
limiter key must include tenant_id — never a process-global singleton.
# Sketch only — see langchain-rate-limits for production implementation.
_limiters: dict[str, InMemoryRateLimiter] = {}
deflimiter_for(tenant_id: str) -> InMemoryRateLimiter:
if tenant_id notin _limiters:
# Tier lookup — different plans get different budgets.
rps = TENANT_TIER_LIMITS.get(tenant_id, 1.0) # 1.0 rps = free-tier default
_limiters[tenant_id] = InMemoryRateLimiter(requests_per_second=rps)
return _limiters[tenant_id]
For multi-process deployments use a Redis-backed limiter — InMemoryRateLimiter
is per-process only, so 1 rps × 8 workers = 8 rps aggregate.
Step 5 — Emit a structured audit log on every invocation
The audit log is the record of truth during an incident. Emit on both success
and failure paths. The common bug is logging in the try block only — when a
tool raises, the except branch re-raises without emitting, so the incident
responder has no record of the denied call.
import json
import time
import uuid
from contextlib import contextmanager
@contextmanagerdefaudit(ctx: dict):
started = time.monotonic()
trace_id = str(uuid.uuid4())
record = {**ctx, "trace_id": trace_id, "outcome": "pending"}
try:
yield record
record["outcome"] = record.get("outcome", "success")
except Exception as exc:
record["outcome"] = "error"
record["error_class"] = type(exc).__name__
record["error_message"] = str(exc)[:500] # truncate — no PII leakageraisefinally:
record["latency_ms"] = int((time.monotonic() - started) * 1000) # 1000 ms per secondprint(json.dumps(record)) # or a SIEM / BigQuery client
ctx = {
"user_id": "u_42",
"tenant_id": "initech",
"chain_name": "rag-qa-v3",
"role": "viewer",
}
with audit(ctx) as record:
result = chain.invoke(inputs, config={"configurable": ctx})
record["tools_called"] = [m.name for m in result.get("tool_calls", [])]
record["input_tokens"] = result["usage"]["input_tokens"]
record["output_tokens"] = result["usage"]["output_tokens"]
record["cost_usd"] = cost_of(result["usage"])
The try / finally pattern guarantees emission even when the chain raises.
Required fields: user_id, tenant_id, chain_name, outcome, latency_ms,
trace_id. Recommended: tools_called, input_tokens, output_tokens,
cost_usd, role. See Audit-log schema for
the full catalog, JSON example, SIEM / BigQuery ingestion, and query recipes.
Step 6 — Two-tenant regression test for cross-tenant leak
The test that would have caught P33 on day one. Seed two tenants with distinct
documents, run a golden query as each, assert non-overlap on retrieved IDs.
import pytest
@pytest.fixturedeftwo_tenant_store():
seed("acme", ["acme-doc-1", "acme-doc-2"])
seed("initech", ["initech-doc-1", "initech-doc-2"])
yield
cleanup(["acme", "initech"])
deftest_tenant_isolation(two_tenant_store):
acme_hits = chain.invoke({"query": "q"}, config={"configurable": {"tenant_id": "acme"}})
initech_hits = chain.invoke({"query": "q"}, config={"configurable": {"tenant_id": "initech"}})
acme_ids = {d.idfor d in acme_hits["documents"]}
initech_ids = {d.idfor d in initech_hits["documents"]}
assert acme_ids.isdisjoint(initech_ids), \
f"CROSS-TENANT LEAK: overlap={acme_ids & initech_ids}"assertall("acme"in doc_id for doc_id in acme_ids)
assertall("initech"in doc_id for doc_id in initech_ids)
Run in CI on every PR. A regression to import-time retriever binding fails
immediately. See
Multi-tenant regression tests
for fixtures covering tool-allowlist violation, audit-log completeness, and
rate-limiter scoping.
Output
Retriever factory keyed by RunnableConfig.configurable["tenant_id"]
Vector-store choice made against the 4-row isolation table, not by defaults
Agent constructed per-request with a role-scoped tool allowlist (not a
deny-all at tool-call time)
Rate limiter and cost budget keyed by tenant_id, never a process global
Structured audit log emitted on both success and failure paths with
required fields (user_id, tenant_id, chain_name, outcome,
latency_ms, trace_id)
Two-tenant pytest fixture in CI that asserts cross-tenant non-overlap
Error Handling
Error / symptom
Cause
Fix
Tenant A's docs returned to Tenant B after deploy
Singleton retriever bound at import (P33)
Move to per-request factory keyed by config["configurable"]["tenant_id"] (Step 1)
KeyError: 'tenant_id' in retriever factory
Caller forgot to pass configurable
Fail fast with PermissionError; never default to a tenant
permission denied for relation documents on PGVector
RLS policy enabled but session variable not set
SET LOCAL app.tenant_id = :tid inside the transaction
Agent calls a tool the user's role should not have
Every tool bound at agent construction
Build agent per-request with only role-permitted tools (Step 3)
Audit log missing entries for errored invocations
Log emitted inside try only, not finally
Move emit to finally block (Step 5)
Shared rate limit trips all tenants when one misbehaves
Process-global InMemoryRateLimiter
Key limiter by tenant_id; use Redis-backed for multi-process (Step 4)
Cross-tenant leak regression ships to prod
No two-tenant CI test
Add the Step 6 fixture; run on every PR
Audit log contains raw PII from prompts
Logging inputs verbatim
Log field names / token counts / IDs only; never log raw message content
Examples
Full multi-tenant RAG agent, end-to-end
Combines all six steps: retriever-per-request, Pinecone namespace isolation,
role-scoped tools, per-tenant rate limit, audit-log context manager, regression
test. Assembly is ~80 lines. See
Retriever-per-request.
Migrating from a shared singleton to per-request
Wrap the old singleton, add the factory alongside, route by feature flag, ship
the regression test first, flip the flag, delete the singleton. Typical
migration window: 1-2 sprints. See
Multi-tenant regression tests.
Exporting audit log to BigQuery for compliance queries
Stream each record to BigQuery per the
Audit-log schema. Recipes: "tool calls by
user X in last 24h", "tenants with >1% error rate", "highest-spend tenants".
