Coming soon · the AI Pack

Your RAG is only as fresh as your last re-index.

Remac keeps your vector database in sync with every row change - in real time. A product description changes in Postgres, the embedding in Pinecone updates seconds later. No cron. No drift. No stale context feeding hallucinations.

Read the docs
the sync pipeline
Postgres row changes
Remac captures via CDC
Embedding transformer → your provider
Vector DB sink writes the embedding
● context current · low sync lag
The RAG problem

Stale context is a data pipeline problem.

Cron & re-index runs
—Embeddings hours or days behind the source—The LLM answers about products that no longer exist—Full re-index burns compute and API budget—Hallucinations, frustration, eroded trust
Remac CDC sync
+Embeddings update within seconds of a change+Context always reflects the live database+Embedding cache skips unchanged content+One YAML pipeline — no cron, no drift
How it works

From row change to fresh embedding — in one config.

The embedding transformer is just another step in the stream. Declare the provider, the input template, and the vector sink - Remac handles capture, embedding, caching, and upsert.

remac.yaml
streams:
  - id: catalog-rag
    source:
      type: postgres
      settings:
        replication_slot: remac_rag
    middlewares:
      filters:
        - type: table
          settings:
            include: [products]
      transformers:
        - type: embedding
          settings:
            provider: openai
            model: text-embedding-3-small
            input: "{{ name }} - {{ description }}"
            cache: true    # skip unchanged
    sinks:
      - type: pinecone
        enabled: true
        settings:
          index: catalog
          namespace: products
1
Capture

CDC decodes the row change off the transaction log (WAL in this case with a Postgres source) — insert, update, or delete.

2
Embed

The transformer renders the input template and calls your provider — cache skips unchanged content.

3
Upsert

The vector lands in your store within seconds — context stays current, no cron.

Vector DB sinks

Every vector database is just a sink.

Same architecture as Kafka or S3 — point the stream at your store and the embedding transformer handles the rest.

QdrantPineconesoonWeaviatesoonMilvussoonpgvectorsoonChromaDBsoon
Embedding providers
OpenAICoheresoonVoyage AIsoonGoogle GeminisoonMistralsoon
Beyond RAG

Fresh data, all the way down the AI stack.

Feature store freshness

Every transaction that touches a feature-relevant table updates the store immediately. Model serving always reads the freshest features — no waiting for the next batch.

Training data pipelines

Continuous archival of production changes to S3 as JSONL, directly queryable by any engine that reads it. Your training pipeline is a YAML file, not an Airflow DAG.

The case for real-time

Cron re-index vs Remac sync.

 
Nightly / cron re-index
Remac AI Pack
Freshness
Hours to days stale between runs
Seconds — per row change
API cost
Re-embeds the whole corpus each run
Only what changed — cache skips the rest
Completeness
Whatever the last run happened to see
Every committed change, in order
Compute profile
Spiky — full re-index hammers the DB
Continuous, low memory footprint
Maintenance
A cron + scripts to babysit
One YAML pipeline
The AI Pack

A bolt-on, not a second product.

Same binary, same pipeline — add it to any paid tier and the AI capabilities switch on. All your regular CDC keeps working, gated by the license, not a separate deploy.

Embedding transformer

Calls OpenAI today; attaches the vector to the event. Cohere, Google Gemini, Mistral, Voyage AI, and more on the roadmap.

Vector DB sinks

Qdrant supported today. Pinecone, Weaviate, Milvus, pgvector, ChromaDB, and more on the roadmap — each just a sink type.

RAG sync monitoring

Embedding latency, write throughput, and sync lag at a glance.

Embedding cache

Skips re-embedding unchanged content — saves real API spend.

Who it’s for

From RAG to training pipelines.

Here’s where teams start. The pipeline doesn’t care what’s downstream.

RAG app builders

Keep the vector store in lockstep with the source so the assistant never cites stale data.

ML platforms

Stream feature-relevant changes to the store so models serve on the freshest features.

Fine-tuning teams

Continuous production-data archival to S3 as JSONL — training sets that build themselves.

AI infra & wrappers

Fan one source out to many stores and services — without owning the pipeline.

AI Pack questions

Which embedding models can I use?+
OpenAI out of the box today. Cohere, Voyage AI, Google Gemini, Mistral, and more are on the roadmap. Set the provider and model in the transformer config.
Does it cache embeddings?+
Yes. With cache: true the transformer skips re-embedding unchanged content, so you only pay the API for rows that actually changed.
What's the sync latency?+
Low latency from source commit to vector upsert. The bottleneck is the embedding API call, not Remac.
Can I keep everything self-hosted?+
Yes — point the transformer at a local model endpoint and use a self-hosted Qdrant as the sink. Nothing has to leave your VPC.
Which vector databases are supported?+
Qdrant is supported today. Pinecone, Weaviate, Milvus, pgvector, ChromaDB, and more are on the roadmap — each will be a sink type, same as Kafka or S3.

Stop re-indexing. Start remacing.

Cron jobs and full re-index runs are a workaround. Real-time CDC is the fix.

See case studies