Skip to main content

E-02 DLP Gateway

Brick E: Data, RAG, context, memory. French version: E-02-dlp-gateway

1. Role and responsibilities​

Inspects and transforms flows (prompts, tool outputs, messages, documents): detect, mask, tokenize, block.

  • Apply eai.dlp_rules
  • Choose the mode by policy and destination
  • Tokenize via E-03 in tokenize mode

2. Position in the architecture and interactions​

  • Callers: Model Gateway, Tool Gateway, Channel Gateway, A2A.
  • Execution path: Synchronous, inline.
  • Every interaction goes through the Control Plane API v1 (/v1) behind the API Gateway (D-14).

3. Open source component and developed part​

  • Open source: Presidio (anonymizer)
  • Developed: Gateway, rules, modes.
  • External integration through an adapter eai/integrations/<system>.py with an in-process fallback when the URL is not configured.

4. Data model​

Tables (schema eai, tenant_id RLS):

  • eai.dlp_rules

Common columns: id uuid, tenant_id uuid NOT NULL, created_at, updated_at; RLS policy current_setting('app.tenant_id', true)::uuid.

5. Contract​

  • Applied contract: policy-contract (obligations dlp)
  • JSON Schema: contracts/v1/policy-contract.schema.json
  • Key fields: obligations[].type=dlp, mode
  • Validation: POST /v1/contracts/{name}/validate; manifest apply: POST /v1/manifests/apply.

6. API endpoints​

  • POST /v1/dlp/inspect
  • POST /v1/dlp/redact
  • GET /v1/dlp/rules
  • POST /v1/dlp/rules

Request POST /v1/dlp/redact:

{
  "text": "Appelez-moi au 06 12 34 56 78",
  "mode": "mask"
}

Response:

{
  "text": "Appelez-moi au <PHONE_NUMBER>",
  "findings": [
    {
      "type": "PHONE_NUMBER"
    }
  ],
  "action": "masked"
}

7. SDK examples​

Python:

from enterprise_ai import EnterpriseAI

client = EnterpriseAI(base_url="<control-plane-url>", api_key="eai_xxx_secret")
out = client.dlp.redact(text="...", mode="mask")

TypeScript:

import { EnterpriseAI } from "@enterprise-ai/sdk";

const client = new EnterpriseAI({ baseUrl: "<control-plane-url>", apiKey: process.env.EAI_API_KEY });
const out = await client.dlp.redact({ text: "...", mode: "mask" });

Java:

EnterpriseAIClient client = EnterpriseAIClient.builder()
    .baseUrl("<control-plane-url>")
    .apiKey(System.getenv("EAI_API_KEY"))
    .build();
JsonNode out = client.dlp().redact(Map.of("text", "Call me at +33 6 12 34 56 78", "mode", "mask"));

Go:

c, _ := eai.NewClient(eai.WithBaseURL("<control-plane-url>"), eai.WithAPIKey(os.Getenv("EAI_API_KEY")))
out, _ := c.DLP.Redact(ctx, map[string]any{"text": "...", "mode": "mask"})

8. Configuration​

  • Environment variables: EAI_DATABASE_URL
  • Helm values keys: controlPlane.config, controlPlane.secrets

9. Security controls​

  • Block mode for secret toward external provider
  • Rules editable by security_officer
  • Every request carries the tenant resolved from the credential; PostgreSQL RLS
  • Every mutation and policy decision writes a hash-chained audit event

10. Observability​

  • Prometheus metrics: eai_dlp_findings_total{type,mode}
  • Traces: OpenTelemetry spans with GenAI semantic conventions (gen_ai.operation.name, gen_ai.agent.id, gen_ai.tool.name), attributes eai.tenant_id, eai.run_id, traceparent propagation.
  • Audit actions: dlp.redact, dlp.inspect

11. FinOps units​

  • Inspected characters.
  • Emitted as Usage Event (POST /v1/finops/usage) with tenant, project, agent and run.

12. Failure modes and resilience​

  • Fail closed on restricted data
  • Normalized errors (error envelope, message_i18n en/fr)

13. Standalone adoption (client of a client)​

  • enterprise_ai.dlp + Control Plane.
  • SDK: enterprise_ai.privacy / dlp / knowledge_bases / memory / retention / lineage
  • Minimal deployment: Helm chart with controlPlane.enabled=true and PostgreSQL; disabled external components use their in-process fallback.

14. Acceptance tests​

  1. block mode -> 403 dlp_blocked
  2. mask replaces the entity
  3. Cross-tenant access refused
  4. Error returned in fr with Accept-Language: fr

Adapter files and streaming output DLP​

  • A file sent to a specialized adapter (for example /v1/gateway/audio/transcriptions) goes through the F11 multimodal DLP before the call: text extracted by type (document, image OCR, audio transcript) and metadata (file name). In block mode the answer is 422 dlp_blocked with direction=file and the adapter is not called.
  • Adapter text responses (text, transcript, ocr_text fields, text/* responses) go through the response DLP with the tenant mode.
  • When streaming, a sliding buffer is always active as soon as the DLP is not off: the end of the text (EAI_STREAM_DLP_HOLDBACK_CHARS, 128 characters by default) is held back, and an entity found across the cut is kept whole for the next chunk. A value split between two chunks is therefore detected and masked.

Governed RAG (kit L4)​

Presidio (presidio-analyzer and presidio-anonymizer, spaCy en_core_web_sm) is tested for real on fictitious banking examples (FR IBAN, Luhn card, French phone, email, name) in tests/test_data_presidio_real.py; the builtin recognizers stay the fallback.