NEW BLOG

Using Surrealism to build your own extensions

1/2

Build context-aware AI agents with SurrealDB

Q: How does SurrealDB support RAG (retrieval-augmented generation)?

SurrealDB stores vector embeddings alongside structured data and supports native HNSW vector indexing. You can combine vector similarity search with graph traversals, filters, and full-text search in a single query to retrieve precisely the context your LLM needs.

Q: Can SurrealDB act as memory for AI agents?

Yes. Agents can persist conversation history, tool outputs, and reasoning chains as structured records with graph relationships. SurrealDB's record links let agents traverse their own memory graph to recall context from previous interactions.

Q: What is hybrid search in SurrealDB?

Hybrid search combines vector similarity with keyword-based full-text search and structured filters in a single SurrealQL query. This gives you more relevant results than vector search alone, especially when domain-specific terms or metadata matter.

Q: Does SurrealDB integrate with LLM frameworks?

Yes. SurrealDB works with popular AI frameworks through its official SDKs for Python, JavaScript, Rust, and others. You can connect it to LangChain, LlamaIndex, or your own orchestration layer to provide agents with persistent storage and retrieval.

Q: Can I run ML models directly inside SurrealDB?

Yes. SurrealML lets you train models in Python using PyTorch, Tensorflow, or Sklearn, then import and run them directly inside SurrealDB. Inference is powered by a Rust-native ONNX runtime, enabling CPU and GPU computation at query time without external service calls.

Give AI agents persistent memory, rich context, and semantic understanding across structured, unstructured, and graph data using a unified, expressive query language.

What AI Agents need to operate reliably

AI agents need a unified foundation to reason, act, and adapt in real time. To perform reliably in production, they require consistent access to context, state, tools, and governance - the core elements that shape scalable agentic systems.

Unified memory

Agents need a single, persistent source of truth for context and history. Fragmented memory across systems leads to lost context and unreliable reasoning.

Fast data access

Agents need sub-millisecond reads and real-time queries to respond within tight latency windows. Slow data retrieval creates bottlenecks that degrade the entire agentic workflow.

Reliable execution

Agents need deterministic, ACID-safe operations so updates and actions remain consistent. Without this, they risk duplicating work or corrupting shared state.

Strong governance

Agents need fine-grained permissions, audit trails, and secure boundaries to safely interact with sensitive or operational data.

THE CHALLENGE

The challenge with legacy backends

Fragmented architectures

Traditional backends rely on multiple disconnected data stores - creating fragile integrations and inconsistent state.

Loss of context

When embeddings, documents, and metadata flow between separate systems, agents lose context, reducing reasoning quality and reliability.

Operational overhead

Legacy stacks require teams to manage many specialised services, increasing maintenance, complexity, and operational risk.

Cost and latency spikes

Multi-system pipelines introduce latency and higher infrastructure costs, especially under vector-heavy or multi-agent workloads.

THE SOLUTION

The solution for AI agents

Unified data and memory

SurrealDB stores structured data, documents, vectors, and graphs in one place, giving agents a single source of context and state.

Multi-model reasoning

A unified query layer lets agents combine relational, graph, and vector operations in one request for faster, more accurate decisions.

Real-time reactivity

Subscriptions and event triggers let agents react instantly to changing data, enabling more adaptive behaviour.

Access and permissions

Role-based access control, permissions, encryption, and auditing provide clear oversight of agent activity, ensuring safer, compliant AI workflows.

CAPABILITIES

Multi-model querying for richer context

SurrealDB enables graphs, vectors, documents, and relational data in one language. Blend connections, facts, and semantics in a single round trip.

1-- Example: Find products similar to a user's last purchase
2
3-- Get the last user's purchase info by datetime
4LET $last_purchase = user:one->purchased.at.last();
5
6-- And the most recently purchased product from that
7LET $last_product = (
8	user:one->purchased[WHERE at = $last_purchase]->product
9)[0];
10
11-- Then get the most similar products
12-- $last_product will show up as most similar so slice starting at index 1
13(
14	SELECT
15		id,
16		vector,
17		vector::similarity::cosine($last_product.vector, vector) AS similarity
18	FROM product
19	ORDER BY similarity DESC
20	LIMIT 3
21)[1..];

1-- Example: Find siblings (i.e., documents that share a tag) and calculate vector (semantic) similarity
2
3SELECT VALUE (
4	SELECT *,
5		vector::similarity::cosine(embedding, $parent.embedding) AS similarity
6	FROM array::distinct(->tagged_with->$tag<-tagged_with<-document)
7) AS siblings
8FROM ONLY $record
9FETCH siblings;

RAG, Graph RAG, and unified memory

Retrieval-Augmented Generation thrives on fresh, structured context. This content is usually stored in memory, but can also be saved to disk for persistent storage. Toolkits in languages including Rust, Python, and JavaScript make this easy.

As an example, let's see what a simple Graph RAG looks like using LangChain. We will use SurrealDB for the vector and graph stores, and Ollama to generate the embeddings.

1# DB connection
2conn = Surreal(url)
3conn.signin({"username": user, "password": password})
4conn.use(ns, db)
5
6# Vector Store
7vector_store = SurrealDBVectorStore(
8	OllamaEmbeddings(model="llama3.2"),
9	conn
10)
11
12# Graph Store
13graph_store = SurrealDBGraph(conn)

1documents.append(Document(page_content=chunk, metadata=...))
2
3# This calculates the embeddings and inserts the documents into the DB
4vector_store.add_documents(documents)

1# Find nodes and edges (Product -> tagged_with -> Tag)
2tag_node = Node(id=tag.name, type="Tag", properties=asdict(tag))
3product_node = Node(id=product.id, type="Product", properties=asdict(product))
4
5# Edges
6relationships.append(
7	Relationship(source=product_node, target=tag_node, type="tagged_with")
8)
9graph_documents.append(
10	GraphDocument(nodes=nodes, relationships=relationships, source=doc)
11)
12
13# Store the graph
14graph_store.add_graph_documents(graph_documents, include_source=True)

SECURITY

Governance and security

Meet regulatory requirements without wrapping your database in yet another proxy.

RBAC & record-based access

On every part of your database schema.

Rate limiting

By disallowing arbitrary queries, requiring low-access users to use defined API endpoints with middleware such as timeouts instead.

Encrypted in flight

With TLS and on-disk encryption.

Audit trails

Use env vars to record logs and traces to file with automatic rotation.

Business impact

Lower total cost of ownership

One tool to replace multiple databases (structured and unstructured, graph, vector), blob store, queue, and function compute.

Reduced complexity

Replace multiple specialised systems for memory, vectors, and relationships with a single, unified SurrealDB platform that's easier to build, operate, and scale.

Stronger governance

Centralised policies and audit across structured, unstructured, and AI data.

Scale as you need

Go from embedded, to single-node server, to a highly scalable distributed cluster.

TRUSTED BY

Powering innovation across industries

Unlocking insights with knowledge graphs

Samsung uses SurrealDB to power a knowledge graph for real-time audience insights and ad targeting in its ad division.

Learn more

Unified infrastructure monitoring

Tencent uses SurrealDB to consolidate nine backend tools into one real-time monitoring platform.

Learn more

Frequently Asked Questions

How does SurrealDB support RAG (retrieval-augmented generation)?

Can SurrealDB act as memory for AI agents?

What is hybrid search in SurrealDB?

Does SurrealDB integrate with LLM frameworks?

Can I run ML models directly inside SurrealDB?

LET'S GOOOO

Ready to build AI agents?

Build dynamic, intelligent systems and enhance AI-driven data analysis for real-time, actionable insights.