Couchbase AI Services

Vector Search, RAG and Productions AI Agents

TechStar Group runs a dedicated Couchbase Center of Excellence — 30+ Couchbase engineers, 25+ certifications — and builds AI systems on the Couchbase platform itself: vector search, hybrid search, retrieval-augmented generation, agent memory, and governed agent access through the Couchbase MCP Server. We design them, build them, and then run them under managed services. Same team, one contract.

No migration is required to start. If you already run Couchbase Server or Couchbase Capella, this is the fastest engagement we offer. The assessment is free and the first proof of concept is free, built in your own environment, with every artifact yours to keep.

Couchbase 8 changed what your data layer can do for AI

Most AI programmes stall in the same place. The pilot works. Then the team tries to put it into production and discovers the architecture is four systems pretending to be one: an operational database, a separate vector store, a cache, and a bolted-on memory layer, each with its own consistency model, its own failure mode and its own bill. Latency compounds, recall degrades, and nobody can explain what the agent did or why.

Couchbase 8, released in October 2025 with over 400 features and changes, removed a large part of that problem by putting vector workloads on the same platform as the operational data. It offers three distinct vector indexing approaches rather than one generalised one — which matters, because AI applications do not have uniform performance and accuracy needs.

The architectural point is simple: vector and AI is not a bolt-on beside your database. It is one service among eight on the same cluster, sharing the same data, the same security model and the same operational runbook.

Couchbase

  1. Mobile / Edge
  2. Analytics
  3. Vector / AI this page
  4. Eventing
  5. Full-Text Search
  6. SQL++ Query
  7. Document
  8. Key-Value
Vector and AI as one of eight Couchbase services on a single platform.

Choosing the right Couchbase 8 vector index

This is the first architectural decision on any Couchbase AI build, and getting it wrong is expensive to unwind. Hyperscale and Composite vector indexes are held by the Index Service; Search Vector indexes are held by the Search Service.

Index typeWhat it is Use it when
Hyperscale Vector Index A single vector column, indexed via global secondary indexing. Built for datasets scaling into the billions of documents, at a low memory footprint. You are primarily querying vector values at very large scale and want speed with low memory cost.
Composite Vector IndexCombines vector and scalar predicates, supporting pre-filtered queries that scope the vectors being searched. Stored and partitioned like other global secondary indexes.You need to narrow the search space first — by tenant, region, category or permission — before the similarity search runs.
Search Vector IndexVector queries through the Search Service, supporting hybrid searches that combine vector, lexical and structured criteria in a single SQL++ request. You need true hybrid search — semantic similarity together with text, geospatial, range or fuzzy matching.

Couchbase 8 also added BM25 best-match scoring for better hybrid search results, read-replica partitions in the Search Service for higher query throughput, and roughly doubled Search Vector Index performance through AVX2 SIMD support. Vector search itself has been available since version 7.6 through the Search Service; version 8.0 is what made the global-secondary-index route available.

The Couchbase AI Data Plane — what production agents actually need

The Couchbase AI Data Plane, formerly Capella AI Services, is where Couchbase put the pieces an agent needs beyond retrieval. It is available as a Capella-managed service and self-managed against Couchbase Server Enterprise Edition. It has three parts worth knowing:

  • Agent Memory

    Short-term, long-term, semantic, profile and conversational memory that survives sessions, restarts, users and framework changes, held next to the operational data rather than in a separate service.

  • Agent Catalog

    Governance for tools and prompts: what tools exist, which agents may use them, what the prompts are, and an end-to-end trace of what actually happened. This is what makes agent behaviour auditable instead of mysterious.

  • Couchbase MCP Server

    Standardised agent access to operational data, vectors, documents, tools, prompts, traces and cache through the Model Context Protocol.

Keeping prompts, tools, traces, memory and operational data in one governed layer has a practical consequence: you can inspect agent behaviour with SQL++, using skills your team already has, rather than stitching together logs from four systems.

Couchbase MCP Server — governed agent access to your data

The Couchbase MCP Server is a self-hosted Model Context Protocol server that lets AI agents and LLM clients interact directly with data in Couchbase clusters, whether on Capella or self-managed. It exposes tools across cluster health, data schema, key-value operations, query and performance, and it supports both STDIO and streamable HTTP transports. It is distributed as a PyPI package and as a Docker image.

The detail that matters for anyone putting this near production data: read-only mode is on by default. Every write operation — key-value, query, scope and collection management, index management — is disabled unless you deliberately enable it, and individual tools can be disabled with fine granularity. Enterprise support for the MCP Server comes with a Couchbase AI Data Plane licence, which also covers Agent Memory and Agent Catalog.

What we do with it: design the tool surface an agent is permitted to see, set the read-only and tool-disabling policy against your data classification, wire it into your agent framework, and then run it — including the part most teams skip, which is watching what the agents actually do once they are live.

The five AI services we deliver on Couchbase

  1. 1

    AI readiness assessment — no charge

    A read-only review of your Couchbase estate against the AI workload you have in mind: whether the cluster is sized for vector workloads, which index strategy fits your query pattern, where your embeddings should live, and what has to change first. You get a defensible effort number before committing budget.

  2. 2

    Vector and hybrid search build

    Embedding strategy and model selection, vectorisation pipeline for structured and unstructured data, index type selection against the table above, index design and tuning, and recall-versus-latency benchmarking on your data rather than on a public dataset. Deployed on-premises, in the cloud, at the edge, or air-gapped.

  3. 3

    Retrieval-augmented generation pipelines

    End-to-end RAG on Couchbase: chunking and ingestion, retrieval design combining semantic and lexical matching, semantic and exact caching to cut redundant LLM calls, grounding and citation handling, and evaluation harnesses so that retrieval quality is measured rather than assumed.

  4. 4

    Agentic systems and MCP

    From single-workflow agents to multi-agent systems: agent memory design, tool and prompt governance through Agent Catalog, MCP Server deployment and policy, tracing and observability, and human-in-the-loop gates where the risk warrants them. We build the agents that handle real operational workflows, not demonstrations.

  5. 5

    Managed AI run

    The layer almost nobody offers. Once it is live, we run it: 24×7 operations, SRE and DBA, upgrades, monitoring, backup and recovery for the Couchbase estate, plus the AI-specific work — index maintenance as data drifts, retrieval-quality monitoring, prompt and tool version control, and cost control on inference.

Why the data layer decides whether your agents reach production

Agentic systems behave like operational applications, not like analytics. A single user interaction can fire many reads, writes, searches and memory lookups, and the traffic pattern is unpredictable as teams add agents. That demands the responsiveness, availability and distributed scale of an operational platform — which is exactly the argument for keeping agent data on the same platform as the operational data rather than in a separate specialised store.

The practical consequences of getting this wrong are consistent across every stalled programme we see: latency that is acceptable in a demo and unacceptable at scale; recall that degrades as the corpus grows because the index type was chosen before the query pattern was understood; inference costs that climb because nothing is cached; and no way to answer the question an auditor or a regulator will eventually ask, which is what the agent saw and why it did what it did.

AI on an estate you have just migrated

The two halves of this practice connect. A migration to Couchbase consolidates a fragmented estate onto one platform; the AI services then run on top of it without a second integration project. Customers who come to us for a migration usually arrive at the AI layer within a year, and customers who come to us for AI often find that the real blocker is an estate still running on four engines.

Read next Couchbase Migration Services Couchbase Managed Services The Couchbase Center of Excellence

Five source families (relational, NoSQL and wide-column, in-memory and grid, cloud-native, search and analytics) converging on one Couchbase destination with eight service layers, including Vector and AI.
Vector and AI sits on the same destination platform every migration lands on.

How to engage us

Read-only review of the estate against your intended AI workload, with index strategy, sizing and a defensible effort number.

No charge

One AI use case built in your own environment — typically a vector search or RAG pipeline on your real data. Scope agreed before we start. Every artifact is yours.

No charge

Full delivery, milestone-gated against evidence.

Fixed price

Managed AI and Couchbase operations, sized to the account.

Monthly

Start with the free AI readiness assessment

A read-only review of your Couchbase estate against the AI workload you are planning, at no charge. If your estate is not ready, we will tell you what has to change before anyone spends money on models.

Book a Couchbase AI readiness assessment
Kishore Tulluri VP, Database as a Service; Head, Couchbase Center of Excellence Krishnakishore.tulluri@techstargroup.com +91 9000 999 054
Raghu Chittimalla Co-founder and Managing Partner, Dallas raghu@techstargroup.com +1 214.402.3125

Couchbase AI services — frequently asked questions

Can Couchbase be used as a vector database?

Yes. Vector search has been available in Couchbase Capella and self-managed Couchbase Server since version 7.6 through the Search Service, and Couchbase 8.0 added vector query through global secondary indexes with two further index types. The practical advantage over a standalone vector database is that the vectors sit alongside the operational data, so you are not running and reconciling two systems.

What is the difference between Hyperscale, Composite and Search Vector indexes?

Hyperscale indexes hold a single vector column and are built for datasets scaling into billions of documents with a low memory footprint. Composite indexes combine vector and scalar predicates so queries can be pre-filtered before the similarity search runs. Search Vector indexes go through the Search Service and support hybrid queries that combine vector, lexical and structured criteria in one SQL++ request. Choose against your query pattern, not against your data volume alone — that is the mistake that is expensive to unwind later.

What is the Couchbase MCP Server?

It is a self-hosted Model Context Protocol server that lets AI agents and LLM clients interact with data in Couchbase clusters, on Capella or self-managed. It provides tools across cluster health, data schema, key-value, query and performance, supports STDIO and streamable HTTP transports, and ships as a PyPI package and a Docker image. Read-only mode is enabled by default and individual tools can be disabled, which is what makes it safe to point at production data.

What is the Couchbase AI Data Plane?

It is the current name for what launched as Capella AI Services. It brings together Agent Memory, Agent Catalog and the Couchbase MCP Server on one platform, available as a Capella-managed service or self-managed against Couchbase Server Enterprise Edition, with enterprise support under licence.

Do we need to migrate to Couchbase before building AI services?

No. If you already run Couchbase Server or Capella, we can start on the AI layer immediately — it is our fastest engagement. Migration only enters the conversation if your data is spread across several platforms and the fragmentation is itself the obstacle.

Can we run this air-gapped or on-premises?

Yes. Couchbase 8 runs on-premises, in the cloud, at the edge and in hybrid deployments, and the AI Data Plane is available self-managed as well as on Capella. We deliver into air-gapped environments where the model must remain inside the customer boundary.

How do you measure whether the RAG pipeline is any good?

With an evaluation harness built against your data and your questions, not a public benchmark. We measure retrieval quality, recall against latency, grounding and citation accuracy, and then re-measure as the corpus drifts. A RAG pipeline that is not evaluated is a demo, not a system.

Who does this work — a general AI team or Couchbase specialists?

The same dedicated Couchbase practice that does the migrations: 30+ Couchbase engineers with 25+ certifications inside a 900-person firm operating since 2002. We do not take work that is not Couchbase, which is the reason the depth exists.