Download Free CE

Features

Everything that ships in SynapCores Community Edition. Items marked Enterprise require an EE license.

Quick summary

SynapCores is a single-binary AI database. Community Edition is free and runs everything below on a single host. Enterprise Edition adds clustering, fine-grained RBAC/SSO, CDC, encryption-at-rest, and SLA-backed support.

  • Graph + Vector + LLM in a single SQL/Cypher query
  • 161 ready-to-run recipes shipped with the binary
  • Bundled local LLM (Llama 3.2 1B GGUF) — no API keys required
  • Multi-provider vision (OpenAI / Anthropic / Ollama LLaVA), configurable in the UI
  • Filesystem Collections — drop a file in a folder, query it as SQL
  • Built in Rust — single binary, no JVM, no Python runtime

1. SQL engine

1.1 SQL surface

  • CREATE / DROP / ALTER / INSERT / UPDATE / DELETE / SELECT
  • JOINs (INNER / LEFT / RIGHT / FULL / CROSS)
  • Subqueries, CTEs (including recursive)
  • UNION / INTERSECT / EXCEPT
  • Window functions (ROW_NUMBER / RANK / LAG / LEAD / ...)
  • GROUP BY ... HAVING, ORDER BY multi-column
  • Stored procedures with control flow + error handling
  • Triggers (BEFORE / AFTER on INSERT / UPDATE / DELETE)
  • Views (regular and materialized)
  • ACID transactions

1.2 Data types

Standard: INTEGER, BIGINT, REAL, DOUBLE, TEXT, VARCHAR, CHAR, BOOLEAN, JSON, JSONB, UUID, TIMESTAMP, DATE, TIME, DECIMAL, BYTEA. AI-native: VECTOR(dimensions), AUDIO, VIDEO, IMAGE, PDF.

1.3 Query optimizer

  • Cost-based optimizer (default)
  • Plan cache with literal-stripped hashing + PG-style hedging
  • Result cache (ON by default)
  • Predicate / projection pushdown
  • Join reordering
  • PK fast-path for point lookups (102× on oltp_point_select)

2. Vector database

2.1 Vector storage + search

-- Cosine distance
embedding <=> query_vector

-- Euclidean distance
embedding <-> query_vector

-- Inner product
embedding <#> query_vector
  • HNSW index (approximate NN, configurable M / ef)
  • Flat index (exact NN for small collections)
  • Cosine / euclidean / inner-product distance metrics
  • Metadata filtering at query time
  • MVCC-aware reads (transactional consistency)

2.2 Embedding generation

-- Bundled MiniLM (no external service)
SELECT EMBED('wireless headphones');

-- Or specify a different model
SELECT EMBED(text, 'minilm');
SELECT EMBED(text, 'bert-base');

MiniLM is bundled and runs in-process. Other embedding models can be plugged via the AI service config (Ollama / OpenAI / Cohere / HuggingFace).

3. Graph database (new in v1.5)

SynapCores v1.5 ships a property-graph engine that runs Cypher inside the same query surface as SQL. You can mix structural pattern matches, vector similarity hops, and LLM scoring in a single query.

3.1 Cypher language

  • MATCH / WHERE / RETURN / ORDER BY / LIMIT
  • MERGE for upsert semantics
  • WITH chaining + UNWIND
  • UNION / UNION ALL
  • Multi-statement bodies (recipe-style setup blocks)
  • CALL <procedure> for graph algorithms

3.2 Hybrid pattern + similarity (SIMILAR_TO)

MATCH (seed:Complaint {id: $id})-[:SIMILAR_TO > 0.85]->(c:Complaint)
WHERE c.id <> seed.id
MATCH (c)-[:DESCRIBES]->(p:Product)<-[:SUPPLIES]-(s:Supplier)
WHERE s.audited_year = 2025
WITH s, p, c, llm_score('rate severity 0..1', c) AS severity
WHERE severity > 0.5
RETURN s.name, p.name, c.id, severity
ORDER BY severity DESC LIMIT 50;

SIMILAR_TO is a synthetic edge type backed by HNSW — you ship vector hops as a regular Cypher relationship. llm_score() is a scalar that pipes the row through a configured LLM and parses the numeric score, usable in WHERE and ORDER BY.

3.3 Graph algorithms (GDS)

  • PageRank
  • Louvain (community detection)
  • Label Propagation
  • Triangle Count
  • Single-source shortest path (more in v1.6)

3.4 Graph CDC + replication

Change data capture from a graph (replicate node / edge mutations out to a downstream sink) is Enterprise.

4. AI Chat + NL2SQL

4.1 Bundled local LLM

CE ships with llama-3.2-1b-instruct-q4_k_m.gguf for in-process inference via llama-cpp. No external API calls, no Ollama daemon, no API keys to start. Drop a different GGUF in the models directory to swap to a larger model (Phi-3-mini, Mistral-7B, etc.).

4.2 External provider plugins

  • OpenAI (GPT-4 / GPT-4o / GPT-4o-mini)
  • Anthropic (Claude 3 / 3.5 Sonnet / Haiku)
  • Ollama (any locally-pulled model)
  • Cohere
  • HuggingFace inference API
  • Native (in-process llama-cpp via GGUF)

4.3 NL2SQL

ASK 'Show me top 5 products by revenue last quarter'
ASK 'Find similar complaints to CX-9001 from suppliers we audited in 2025'

Schema-aware: the planner injects the live table catalog into the LLM prompt. Falls back to a deterministic pattern matcher when no LLM is configured.

5. Vision / multimodal (new in v1.5)

Filesystem Collections and Media Gallery use a vision-capable LLM to describe images you ingest. The provider is configured from the Settings UI (or via env vars):

  • OpenAI GPT-4o / GPT-4o-mini (~$0.001 / image)
  • Anthropic Claude 3.5 Sonnet / Haiku
  • Local Ollama LLaVA (free, slower)
  • OCR (Tesseract bundled — works zero-config)

API key encrypted at rest with AES-256-GCM (key derived from AIDB_JWT_SECRET via HKDF-SHA256). Settings live at <data_dir>/system/vision.json, mode 0600.

6. Filesystem Collections

Drop a PDF, image, CSV, audio, or video into a watched folder and SynapCores ingests it automatically: text gets chunked + embedded, images get captioned + OCR'd, audio/video get transcribed (when a Whisper model is on the host).

  • Drop-a-file-into-a-folder ingest pipeline
  • PDF / Office / Markdown text extraction
  • OCR for screenshots and scanned docs
  • Image captioning via configured vision provider
  • Whisper transcription for audio and video (drop the model file)
  • Live progress events via WebSocket
  • Re-process individual documents
  • CREATE COLLECTION SQL syntax

7. Recipes + AutoML

7.1 Recipe library

161 ready-to-run recipes ship inside the binary, organized by category (graph, core foundations, image management, document processing, ML hello-world, security & surveillance, ...). Browse + execute from the bundled web UI.

7.2 AutoML

CREATE EXPERIMENT churn AS
SELECT customer_id, age, tenure, monthly_charges, churned
FROM customers
WITH (
    task_type='binary_classification',
    target_column='churned',
    max_trials=50,
    algorithms=['random_forest', 'xgboost', 'neural_network']
);

DEPLOY MODEL best_model FROM EXPERIMENT churn;
PREDICT churn_probability USING churn_model
  AS SELECT * FROM new_customers;
  • Classification (binary + multi-class)
  • Regression
  • Time-series forecasting
  • Clustering
  • Anomaly detection
  • Automated feature engineering
  • Hyperparameter sweeps

8. Performance

Community Edition ships ten waves of optimizer + storage work as on-by-default behavior:

  • Unified optimizer with PG-style plan-cache hedging (default ON)
  • Result cache (default ON)
  • Binary big-endian PK keys for point-select fast path
  • Bincode V1 doc storage (replaces JSON)
  • RocksDB range_iter for efficient range scans
  • Audit-log O(1) hot path (no more O(n) per query)
  • Lock-free executor reads (ArcSwap)
  • 102× improvement on oltp_point_select vs. v1.4 baseline

Reproducible benchmark harness ships in aidb-bench; PG / MariaDB / MySQL comparison numbers are published in the release notes.

9. Multi-tenancy

Tenant isolation is Enterprise. CE binaries actively license-lock the tenant_isolation.enabled config flag — setting it in CE has no effect (ignored at config-load with a warning). Single-tenant deployments — laptops, single VMs, dedicated single-team hosts — are CE's sweet spot.

10. Clustering / replication / CDC

Multi-node clustering (Raft consensus + replication) and inbound CDC from MySQL/Postgres binlog are Enterprise. CE is single-host by design. Backup and restore are in CE; scheduled backups + PITR are Enterprise.

11. Security

  • JWT auth (CE)
  • API key auth (CE)
  • Argon2 password hashing (CE)
  • TLS via rustls (CE)
  • Encrypted vision-provider keys (CE — AES-256-GCM)

Fine-grained RBAC, SSO/SAML, LDAP, encryption-at-rest for the whole datastore, and audit-unlimited are Enterprise.

12. Developer experience

  • Single binary — no JVM, no Python runtime
  • REST API + WebSocket + native socket protocol
  • Bundled React web UI on the same port
  • Bundled docs (161 recipes + reference)
  • Cypher and SQL on the same connection
  • Prepared statements with $1 / $2 PG-style placeholders

Where to next

Download CE to try everything above on your laptop. For clustering, RBAC/SSO, or paid SLA-backed support, contact sales.