Redis Vector Similarity & Embedding Drift Sentinel

id: redis-vector-similarity-and-embedding-drift-sentinel
namespace: company.ai

inputs:
  - id: redis_url
    type: STRING
    defaults: "redis://localhost:6379"
    description: Connection URL for the Redis instance (e.g.
      redis://:password@host:6379 or redis://localhost:6379).

  - id: vector_key
    type: STRING
    defaults: "kb_embeddings"
    description: Key of the Redis vector set to audit for nearest-neighbor recall
      and similarity health.

  - id: canary_vector
    type: ARRAY
    itemType: FLOAT
    defaults:
      - 0.12
      - -0.34
      - 0.56
      - 0.78
    description: Golden reference embedding vector used to probe nearest-neighbor
      recall in the target vector set.

  - id: min_similarity_threshold
    type: FLOAT
    defaults: 0.65
    description: Minimum acceptable similarity score for the top match before
      flagging semantic drift.

  - id: search_count
    type: INT
    defaults: 5
    description: Number of nearest-neighbor candidates to retrieve from the Redis
      vector set during the probe.

triggers:
  - id: periodic_schedule
    type: io.kestra.plugin.core.trigger.Schedule
    cron: "0 */3 * * *"
    disabled: true
    inputs:
      redis_url: "redis://localhost:6379"
      vector_key: "kb_embeddings"
      canary_vector:
        - 0.12
        - -0.34
        - 0.56
        - 0.78
      min_similarity_threshold: 0.65
      search_count: 5

tasks:
  - id: probe_canary_similarity
    type: io.kestra.plugin.redis.vector.Similarity
    url: "{{ inputs.redis_url }}"
    key: "{{ inputs.vector_key }}"
    vector: "{{ inputs.canary_vector }}"
    count: "{{ inputs.search_count }}"
    allowFailure: true
    description: Executes a VSIM similarity search against the Redis vector set
      using a golden reference embedding to measure recall and distance.

  - id: classify_vector_posture
    type: io.kestra.plugin.core.debug.Return
    format: >-
      {{
        (outputs.probe_canary_similarity is not defined or outputs.probe_canary_similarity.matches is not defined) ? 'REDIS_CONNECTION_FAILED' :
        (outputs.probe_canary_similarity.matches | length == 0) ? 'ZERO_HIT_DRIFT' :
        (outputs.probe_canary_similarity.scores[outputs.probe_canary_similarity.matches | first] < inputs.min_similarity_threshold) ? 'LOW_SIMILARITY_SCORE' :
        'HEALTHY'
      }}
    description: Evaluates nearest-neighbor match counts and top similarity scores
      against operational thresholds.

  - id: route_posture_action
    type: io.kestra.plugin.core.flow.Switch
    value: "{{ outputs.classify_vector_posture.value }}"
    cases:
      ZERO_HIT_DRIFT:
        - id: alert_zero_hit_drift
          type: io.kestra.plugin.slack.notifications.SlackIncomingWebhook
          url: "{{ secret('SLACK_WEBHOOK_URL') }}"
          payload: "{{ {'text': '🚨 *CRITICAL: Redis Vector Zero-Hit Drift Detected*

            • *Vector Set:* `' ~ inputs.vector_key ~ '`

            • *Endpoint:* `' ~ inputs.redis_url ~ '`

            • *Status:* Golden reference query returned 0 matches.

            • *Execution ID:* `' ~ execution.id ~ '`

            *Action Required:* Vector index is empty or HNSW graph indexing has
            corrupted.'} | toJson }}"

      LOW_SIMILARITY_SCORE:
        - id: alert_low_similarity
          type: io.kestra.plugin.slack.notifications.SlackIncomingWebhook
          url: "{{ secret('SLACK_WEBHOOK_URL') }}"
          payload: "{{ {'text': '⚠️ *WARNING: Redis Vector Semantic Drift*

            • *Vector Set:* `' ~ inputs.vector_key ~ '`

            • *Top Match ID:* `' ~ (outputs.probe_canary_similarity.matches |
            first) ~ '`

            • *Score:* ' ~
            outputs.probe_canary_similarity.scores[outputs.probe_canary_similar\
            ity.matches | first] ~ ' (Threshold: ' ~
            inputs.min_similarity_threshold ~ ')

            • *Execution ID:* `' ~ execution.id ~ '`

            *Action Required:* Embedding model drift or distance metric mismatch
            detected.'} | toJson }}"

      REDIS_CONNECTION_FAILED:
        - id: alert_connection_failure
          type: io.kestra.plugin.slack.notifications.SlackIncomingWebhook
          url: "{{ secret('SLACK_WEBHOOK_URL') }}"
          payload: "{{ {'text': '🔥 *ALERT: Redis Vector Store Unreachable*

            • *Endpoint:* `' ~ inputs.redis_url ~ '`

            • *Vector Key:* `' ~ inputs.vector_key ~ '`

            • *Error:* Unable to establish connection to Redis server.

            • *Execution ID:* `' ~ execution.id ~ '`'} | toJson }}"

      HEALTHY:
        - id: log_vector_healthy
          type: io.kestra.plugin.core.log.Log
          message: "Redis vector set '{{ inputs.vector_key }}' is healthy. Top match '{{
            outputs.probe_canary_similarity.matches | first }}' scored {{
            outputs.probe_canary_similarity.scores[outputs.probe_canary_similar\
            ity.matches | first] }} (matches: {{
            outputs.probe_canary_similarity.matches | length }})."
    description: Routes Slack incident notifications or diagnostic logs depending on
      the evaluated vector health.

  - id: audit_sentinel_run
    type: io.kestra.plugin.core.log.Log
    message: "Redis vector sentinel cycle complete for execution {{ execution.id }}.
      Verdict: {{ outputs.classify_vector_posture.value }}, Top score: {{
      outputs.probe_canary_similarity.scores[outputs.probe_canary_similarity.ma\
      tches | first] ?? 'N/A' }}, Total matches: {{
      outputs.probe_canary_similarity.matches | length ?? 0 }}."
    description: Emits an immutable audit log entry capturing vector retrieval
      scores and cycle verdicts.

errors:
  - id: alert_flow_error
    type: io.kestra.plugin.slack.notifications.SlackIncomingWebhook
    url: "{{ secret('SLACK_WEBHOOK_URL') }}"
    payload: "{{ {'text': '❌ *Redis Vector Sentinel Execution Error*

      Flow failed during execution `' ~ execution.id ~ '` against `' ~
      inputs.redis_url ~ '`. Inspect Kestra run logs for details.'} | toJson }}"

Production AI architectures relying on Redis vector sets for real-time Retrieval-Augmented Generation (RAG), semantic caching, and recommendation indexing frequently suffer from silent retrieval failures. When upstream embedding models drift, dimensions shift, or memory eviction drops index nodes, vector similarity scores silently deteriorate. Applications receive degraded context or zero matches without generating traditional database errors, quietly breaking downstream LLM generation quality.

This blueprint implements an automated Day 2 operations sentinel for Redis vector sets. On a scheduled interval, it executes a native VSIM similarity search using a golden reference embedding, evaluates top match scores against strict SLA thresholds, detects zero-hit and semantic degradation anomalies, and dispatches real-time Slack notifications before retrieval issues impact end-user AI applications.

How it works

The periodic_schedule trigger runs the flow on a 3-hour cadence with explicit parameter bindings, shipped in a disabled state for initial validation.
probe_canary_similarity (io.kestra.plugin.redis.vector.Similarity) executes a VSIM command against vector_key using canary_vector, querying for the top nearest neighbors with similarity scoring.
classify_vector_posture (io.kestra.plugin.core.debug.Return) inspects the returned matches and scores map using Pebble logic to determine operational health: REDIS_CONNECTION_FAILED, ZERO_HIT_DRIFT, LOW_SIMILARITY_SCORE, or HEALTHY.
route_posture_action (io.kestra.plugin.core.flow.Switch) directs subsequent steps: zero-hit outcomes and low similarity scores fire high-urgency Slack alerts with element IDs and distance scores, while healthy runs emit diagnostic log metrics.
audit_sentinel_run (io.kestra.plugin.core.log.Log) writes an immutable summary entry capturing the execution ID, evaluated verdict, and top score for compliance auditing.
A flow-level errors handler catches uncaught Redis connection drops or authentication timeouts and dispatches an incident alert to Slack.

What you get

Automated, continuous validation of Redis vector set recall and nearest-neighbor retrieval.
Early warning detection of embedding drift, index corruption, and zero-match search anomalies.
Strict SLA verification on similarity distance thresholds before retrieval quality impacts production RAG.
Actionable Slack notifications pinpointing vector keys, top match IDs, and observed similarity scores.
Full execution audit history tracking semantic similarity trends over time.

Who it's for

MLOps and AI Engineers operating real-time RAG pipelines and vector search on Redis.
Platform Engineers and SREs monitoring in-memory cache and vector database reliability.
Data Engineers maintaining semantic embedding indexes and document search systems.

Why orchestrate this with Kestra

Monitoring vector databases with custom Python cron scripts introduces unmanaged infrastructure, plain-text credential leaks, and fragmented error tracking. Kestra provides native Redis vector tasks built on Lettuce, declarative scheduling, centralized secret management, granular conditional routing, and automated multi-channel alerting without writing or maintaining dedicated monitoring containers.

Prerequisites

A running Redis instance (version 7.2+ or Redis Stack) with vector capabilities enabled, reachable from the Kestra worker environment.
A populated Redis vector set with indexed embeddings (e.g. kb_embeddings).
An incoming Slack Webhook URL configured in Kestra secrets.

Secrets

SLACK_WEBHOOK_URL: The destination Slack webhook URL for operational incident dispatching.
REDIS_PASSWORD: (Optional) Redis authentication password if required by the target cluster.

Quick start

Configure SLACK_WEBHOOK_URL in your Kestra namespace secrets.

Start a local Redis Stack container supporting vector sets:

docker run -d --name redis-vector-sentinel \
  -p 6379:6379 \
  redis/redis-stack-server:latest

Connect the Redis container to your Kestra Docker network:

docker network connect YOUR_KESTRA_NETWORK redis-vector-sentinel

Seed a sample vector into the Redis vector set using Redis CLI or io.kestra.plugin.redis.vector.Add:

docker exec -it redis-vector-sentinel redis-cli VADD kb_embeddings 0.12 -0.34 0.56 0.78 doc_canary

Execute a manual test run in the Kestra UI and confirm that HEALTHY is classified and logged.
Enable the periodic_schedule trigger to activate autonomous 3-hour background monitoring.

How to extend

Add automated re-indexing by chaining io.kestra.plugin.redis.vector.Add to reload fresh golden embeddings when ZERO_HIT_DRIFT is encountered.
Incorporate multi-vector sweeps by wrapping the probe in an io.kestra.plugin.core.flow.Loop to evaluate canary embeddings across different document domains in a single execution.
Route critical alerts to PagerDuty via io.kestra.plugin.pagerduty when similarity scores persistently drop below acceptable thresholds.

Redis Vector Similarity and Embedding Drift Sentinel

How it works

What you get

Who it's for

Why orchestrate this with Kestra

Prerequisites

Secrets

Quick start

How to extend

Links

Tasks

Orchestrate with Kestra

Share this Blueprint

Create a Blueprint

New to Kestra?

Redis Vector Similarity and Embedding Drift Sentinel

How it works

What you get

Who it's for

Why orchestrate this with Kestra

Prerequisites

Secrets

Quick start

How to extend

Links

Tasks

Orchestrate with Kestra

Share this Blueprint

Create a Blueprint

More AI Blueprints

Automated RAG Vector Index Drift and Hallucination Circuit Breaker

Measure a Presidio PII scrub, then block the load when residual PII survives

Audit Qdrant Vector Collection Indexing Health and Status

RAG Synthetic Golden Dataset Generator for LLM Evaluation

Promptfoo LLM Red-Teaming and Security Regression Gate

Detect Qdrant embedding drift and verify bounded re-embedding

New to Kestra?