Capabilities & how to enable each¶
SynapCores is one binary that bundles a lot: SQL, vectors, an in-process LLM, decisions, agents, graph, ML, audit storage, and a MySQL wire. This page is the switchboard — for each capability: what it is, the one step to turn it on, and the one call to use it. Most are on out of the box; a few need a one-time pull or flag.
First time here? Install the engine with the Quickstart, wire a model with First-run AI setup, then come back and switch on what you need.
How 'enable' works
Everything AI-shaped reads one config block, [query.ai_service] (see
AI setup). "Enable" below means one of: already on, pull a model once
(synapcores pull <name>), or set one config/env flag and restart. Authenticate every
call with a JWT (POST /v1/auth/login) or an aidb_… API key as
Authorization: Bearer ….
At a glance¶
| Capability | Enable | Use it with |
|---|---|---|
| SQL + transactions | ✅ on | POST /v1/query/execute |
| Vector search | ✅ on (embedding model) | EMBED(), COSINE_SIMILARITY(), VECTOR(N) |
| Text generation | chat model | GENERATE() |
| AI Chat + NL2SQL | chat model | POST /v1/ai/chat |
| In-database agents | tool-capable chat model | AGENT_RUN(), CREATE AGENT |
| Agent memory | embedding model | MEMORY_STORE / MEMORY_RECALL |
| One-call decisions (d1) | pull d1:latest |
POST /v1/systemone |
| Claim verification (NLI) | --with-nli / AIDB_NLI_ENABLED=1 |
NLI_VERIFY(), /v1/ai/nli/verify |
| AutoML / in-database ML | ✅ on | CREATE EXPERIMENT → AUTOML.PREDICT() |
| Graph / Cypher | ✅ on | POST /v1/graph/match |
| Native vision | vision-capable model | POST /v1/multimodal/describe |
| Filesystem RAG | ✅ on | CREATE COLLECTION |
| Immutable audit tables | ✅ on (key for encryption) | CREATE IMMUTABLE TABLE |
| Data lake export | lake destination + grant | COPY … TO LAKE |
| MySQL wire protocol | [mysql_wire] enabled = true |
any MySQL client, port 3307 |
| MCP (LLM clients) | bridge config | Claude / Cursor / Continue / … |
SQL + transactions¶
What — a full SQL engine with durable, isolated transactions (BEGIN / COMMIT /
SAVEPOINT), enforced FOREIGN KEY / UNIQUE, and MySQL-compatible syntax.
Enable — on by default.
Use — every statement runs through POST /v1/query/execute with {"sql": "..."}:
curl -sS -H "Authorization: Bearer $TOKEN" -H 'content-type: application/json' \
-d '{"sql":"CREATE TABLE t (id INT PRIMARY KEY, name TEXT); INSERT INTO t VALUES (1,'\''a'\'')"}' \
http://localhost:8080/v1/query/execute
See the SQL reference for the full dialect.
Vector search¶
What — semantic search in plain SQL: a VECTOR(N) column, EMBED() to turn text into a
vector, COSINE_SIMILARITY() to rank, and an HNSW index for scale.
Enable — on by default; embedding_model (default all-minilm:latest, 384-dim) pulls
on first use.
Use:
CREATE TABLE docs (id INT PRIMARY KEY, body TEXT, embedding VECTOR(384));
INSERT INTO docs VALUES (1, 'Refunds take 30 days', EMBED('Refunds take 30 days'));
SELECT id, COSINE_SIMILARITY(embedding, EMBED('how long for a refund?')) AS sim
FROM docs ORDER BY sim DESC LIMIT 5;
-- at scale: CREATE INDEX docs_hnsw ON docs(embedding) USING HNSW;
Declare VECTOR(N) to match your embedding model's width (384 native, 1536 OpenAI small).
Text generation¶
What — call an LLM as a SQL value.
Enable — a chat model wired in AI setup (default qwen2.5:7b).
Use — GENERATE('prompt'), optionally GENERATE(prompt, json_object('temperature', 0.2)):
AI Chat + NL2SQL¶
What — a grounded chat assistant over your schema, and natural-language → SQL.
Enable — a chat model (as above).
Use — POST /v1/ai/chat (and POST /v1/ai/chat/stream for SSE), or the AI Chat tab
in the Web UI. NL2SQL is also exposed at POST /v1/nl2sql/query.
In-database agents¶
What — a ReAct agent loop callable from SQL (AGENT_RUN), and durable agents
(CREATE AGENT) that fire on a schedule or on committed DML.
Enable — a tool-capable chat model (qwen2.5:7b qualifies). The in-process local
provider does native tool calling — no Ollama required.
Use:
SELECT AGENT_RUN('aidb-assistant',
'Find the 3 tables with the most rows and explain what each is for') AS report;
AGENT_RUN(persona, task [, json_object('model', …, 'max_iterations', …, 'timeout_ms', …)]).
Durable agents and their tamper-evident run log (_system_agent_runs) arrived in v1.9.0+.
Agent memory¶
What — persistent, semantic memory for chatbots and agents with no schema of your own.
Enable — an embedding model (as for vector search).
Use — MEMORY_STORE(ns, content) to write, and MEMORY_RECALL (table-valued, goes in
FROM) to retrieve:
SELECT MEMORY_STORE('conv_42', 'The user prefers email over phone');
SELECT content, similarity FROM MEMORY_RECALL('conv_42', 'how should I contact them?', 3);
Also MEMORY_UPSERT (revise/retract a belief) and MEMORY_FORGET.
One-call decisions (d1)¶
What — a typed decision (routing choice, yes/no probability, or urgency score) in one call with zero generated tokens, backed by the purpose-built d1 model. Enable — pull d1 once:
Use — POST /v1/systemone, naming "model": "d1:latest" per request. Full guide:
Decision model (d1).
Claim verification (NLI)¶
What — an in-process cross-encoder that verifies a statement against trusted evidence and
returns TRUTH / LIE / UNVERIFIED_PLAUSIBLE. The go-to hallucination gate.
Enable — one-time opt-in: ./install-ce.sh --with-nli (or AIDB_NLI_ENABLED=1 in Docker),
which pulls nli-default (~576 MB) and enables it on the next start.
Use — NLI_VERIFY() / NLI_CLASSIFY() in SQL, or POST /v1/ai/nli/verify. Full guide:
Claim verification (NLI).
AutoML / in-database ML¶
What — train, deploy, and predict with classical ML (classification, regression, clustering, forecasting) directly in SQL — no external ML stack. Enable — on by default (CE caps: 5 models/tenant, 1M training rows/run — see Limits). Use:
CREATE EXPERIMENT churn_model AS
SELECT tenure, monthly_charges, churned AS target FROM customers
WITH (task_type = 'classification', target_column = 'target');
SELECT AUTOML.PREDICT('churn_model', tenure, monthly_charges) AS churn_risk
FROM new_customers;
Graph / Cypher¶
What — a property graph co-located with your tables: MATCH / MERGE / CALL <algo>,
indexed, no ETL.
Enable — on by default.
Use — POST /v1/graph/match with the Cypher in a sql field:
curl -sS -H "Authorization: Bearer $TOKEN" -H 'content-type: application/json' \
-d '{"sql":"MATCH (a:Person)-[:KNOWS]->(b:Person) RETURN a.name, b.name LIMIT 10"}' \
http://localhost:8080/v1/graph/match
Graph queries need an authenticated tenant context (send a JWT or API key).
Native vision¶
What — describe/interpret an image with a vision LLM entirely in-process (native
LLaVA) — no cloud, no API key, the image never leaves the box.
Enable — a vision-capable model configured on the AI service.
Use — POST /v1/multimodal/describe with {"image": "<base64>", "prompt": "..."}. See the
v1.14.0-ce notes.
Filesystem RAG¶
What — point a folder at the engine and get a watched, chunked, embedded collection; CSVs
also become queryable tables.
Enable — on by default.
Use — CREATE COLLECTION from SQL or the Web UI. Full guide:
Filesystem Collections.
Immutable audit tables¶
What — append-only, hash-chained tables for tamper-evident audit trails; UPDATE /
DELETE are rejected, and VERIFY TABLE proves the chain.
Enable — on by default. Optional encryption-at-rest needs a key:
export AIDB_IMMUTABLE_MASTER_KEY=$(openssl rand -hex 32) (hex, ≥ 32 bytes) — required only
when a table requests ENCRYPTION.
Use:
CREATE IMMUTABLE TABLE audit_log (id BIGINT PRIMARY KEY, actor TEXT, action TEXT)
WITH (ENCRYPTION = 'AES256GCM'); -- omit WITH for an unencrypted audit chain
VERIFY TABLE audit_log; -- attests the hash chain end to end
Data lake export¶
What — materialize any SELECT / WITH result as Hive-partitioned Parquet in your object
store, in one statement.
Enable — a configured lake destination with a write_allowlist grant, run by a
full-data-write principal (see the authorization notes in the
v2.4.0-ce release page).
Use — COPY (<query>) TO LAKE '<path>' WITH (CONNECTION = '<dest>', PARTITION_BY = (col)).
MySQL wire protocol¶
What — connect Metabase, DBeaver, DataGrip, or any MySQL client with just a connection
string.
Enable — add a [mysql_wire] block and restart (off by default):
Use — mysql -h <host> -P 3307 -u admin -p. See the
v1.11.0-ce notes.
MCP (for LLM clients)¶
What — expose the engine's SQL/tools to Claude, Cursor, Continue, Windsurf, Zed, and LangChain over the Model Context Protocol, so an LLM can query your database directly. Enable — configure the MCP bridge in your client. Use — full walkthrough: MCP setup.
Where next¶
- First-run AI setup — wire the provider the AI capabilities share.
- SQL reference — every function and statement, version-matched.
- Recipes — 160+ runnable end-to-end examples.
- Community vs Enterprise · Limits & quotas