synesis-graph: Universal Graph Pipeline

Transform your qualitative research into a living, navigable Knowledge Graph ready for AI (GraphRAG).

Language: English | Portugues

Documentation: Synesis Language Docs

This repository contains the official pipeline from the Synesis language to graph representations. It acts as a bridge between structured human analysis (.syn files) and computational intelligence, enabling MCP Agents and Data Science algorithms to interact with your research.

Three backends ship today — Neo4j and ArcadeDB (property graphs) and HTML (self-contained interactive visualization). All are built on the same BackendAdapter contract, so support for further graph databases can be added without touching the pipeline.

Choosing a backend

All three database backends produce the same graph from the same project. On the FACE/UFMG corpus every structural count matches between Neo4j and ArcadeDB; on the Quinto_Andar corpus (41,474 items, 7,293 concepts, 661 sources) the embedded backend matches the Neo4j Aura export on all 16 node and relationship counts. They differ in how the advanced metrics are scoped — see Graph Metrics.

arcadedb or arcadedb-embedded? Same engine, different delivery. Use the server backend when several people share one graph, or when something remote has to reach it. Use the embedded one to work alone: there is nothing to install beyond pip, nothing to start before working, and the graph is a directory you can copy or delete. Phase B below turns that directory into something a chat client can query, still without a server to administer.

Highlights

• Zero-IO / Direct Link: Uses the synesis.load() API to compile the project in memory and sync directly to the database. No intermediate JSON/CSV files.
• Universal & Agnostic: Does not depend on fixed rules (like "Factors" or "Dimensions"). The script reads your Template (.synt) and creates the graph structure dynamically.
• Automatic Metrics: Calculates network metrics at two levels:
  • Native Metrics: Always available via pure Cypher.
  • GDS Metrics: Advanced algorithms when the Graph Data Science plugin is installed.
• Self-Describing Graphs: Every sync writes a ProjectContext vertex carrying your template's semantics — field descriptions, value scales and GUIDELINES — so any consumer of the graph, from Claude Desktop to any MCP client, knows what the data means.
• Full Traceability: Every node and edge maintains origin metadata (source_file, line, column), ensuring scientific auditability.

Architecture

The pipeline follows the "Research as Code" flow:

1. Input: Plain text files (.syn, .syno, .synt, .synp) defined by the Synesis language.
2. Compilation: The compiler validates syntax and semantics in real-time.
3. Dynamic Modeling: The script translates Template definitions into graph structures.
4. Persistence: Data is injected into Neo4j via atomic transactions.
5. Analytics: Graph metrics are calculated and stored in nodes.

Data Modeling: Template → Graph

The pipeline automatically translates field types defined in your Template (.synt) to structures in the Neo4j graph. This translation is dynamic and does not depend on specific field names.

Mapping Table

Base Nodes (Always Created)

Translation Example

Template:

FIELD category TYPE CODE
    SCOPE ITEM
END FIELD

FIELD theme TYPE TOPIC
    SCOPE ONTOLOGY
END FIELD

Resulting Graph:

(:Item)-[:MENTIONS]->(:Category)-[:GROUPED_BY]->(:Theme)

Self-Describing Graphs: ProjectContext

Every sync to a database backend writes a single ProjectContext vertex carrying the project's own context. The context travels with the data, not with the tool.

Without it, an exported graph is syntax without semantics. A consumer that introspects the schema learns that a vertex Aspect has a property name — but not that Aspect is Dooyeweerd's modal scale, that its values are ordered, or what [15] Fiducial means. All of that is declared in your template, and used to be discarded at export time.

What it holds

Why it matters

It answers questions no schema can. The GUIDELINES you wrote for each field are your coding protocol — the explicit decision rules and examples that say what counts as a valid instance of a field. An assistant reading them stops inferring your criteria from the data and starts knowing them.

It is derived, never hardcoded. Two projects with different templates produce different documents: one may have aspect on a 16-value philosophical scale, another categoria_resultado with its own vocabulary. Nothing here assumes one project's words.

It cannot silently drift. Both backends clear the graph before syncing, so the context always describes that snapshot. generated_at lets a consumer say "this graph is three months old" instead of presenting stale data as current.

It is Markdown, not JSON. Measured through a real MCP client: as JSON the field specs arrived as ~7.3k tokens with 53% of the keys null, and the GUIDELINES — which you wrote with headings and line breaks — came escaped inside a string. Prose is smaller, needs no parsing, and keeps the shape you gave it.

Reading it

// Cheap orientation: what is this graph about, and how big?
MATCH (p:ProjectContext) RETURN p.description, p.project_summary

// The full template semantics, when a question needs them
MATCH (p:ProjectContext) RETURN p.template_doc

Backends: ArcadeDB and Neo4j. The HTML backend is out of scope — it is a visualisation artifact with no programmatic consumer, and for a human reading it on screen the context is already implicit.

A graph produced by an earlier version simply has no such vertex; consumers should treat its absence as "no context available", not as an error.

Graph Metrics

The pipeline automatically calculates network metrics that enrich the analysis. Metrics are divided into two levels:

Native Metrics (Always Available)

Calculated via pure Cypher, without external dependencies.

Concept Nodes (CODE)

Taxonomy Nodes (TOPIC, ASPECT, DIMENSION)

Source Nodes

Advanced Metrics

The same three metrics are produced by both database backends, from different engines:

The scores are not directly comparable between the two. Neo4j's GDS projects only the concept subgraph (concepts and their RELATES_TO edges), while ArcadeDB's algo.* procedures run over the whole graph and accept no scope filter — so MENTIONS, GROUPED_BY and the other edges contribute as well. Concepts cited by many items rank higher under ArcadeDB. On the FACE/UFMG corpus the two top-10 PageRank lists overlap by 6 of 10 entries. Both are valid centralities; they answer slightly different questions.

Values are written only onto concept nodes in both cases. The pipeline prints this caveat on every ArcadeDB export.

Neo4j (GDS)

When the Neo4j Graph Data Science plugin is installed, the pipeline calculates advanced network metrics:

Graph Projection Strategies

GDS metrics calculation automatically adapts to template type:

Note: If GDS is not installed, the pipeline displays a warning and continues normally with native metrics.

ArcadeDB (built-in)

ArcadeDB ships a native algorithm library, so no plugin is involved and there is no degraded path: pagerank, betweenness and community are always computed. Graph projection strategies do not apply — the algorithms always see the full graph.

Semantic Search with Vector Embeddings (ArcadeDB)

Full-text search finds concepts that share words with the question. Vector search finds concepts that mean what the question means. The two are complementary, and ArcadeDB indexes both on the same node.

pip install "synesis-graph[embeddings]"

synesis-graph arcadedb --project ./analysis.synp \
    --vector-embeddings ontology_description,topic

What it buys you

Measured on the FACE/UFMG corpus (210 concepts, Portuguese), with five questions whose vocabulary is deliberately disjoint from the concept descriptions:

The vector search returns the exact concept in 4 of 5; BM25 in none. This is the GraphRAG case: the researcher asks in their own words, not in the ontology's controlled vocabulary. Keep the full-text index — it still wins when the term is present.

Which fields to embed

Fields are validated against your template. The rule follows the declared TYPE:

The concept name is always included. Not every TEXT field is a good choice: a field holding the reason for a classification describes the criterion, not the concept, and bipolar fields (Low_Cost/High_Cost) put antonyms in one vector.

Choosing a model

Per project, in config.toml — a Portuguese corpus and an English one have different needs:

[arcadedb.embeddings]
model = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
fields = ["ontology_description", "topic"]

Multilingual is a requirement, not a preference, for Portuguese corpora. On the question above most dependent on Portuguese semantics, the English-only model returned jogo_de_empresa — reproducing exactly the lexical error BM25 makes.

Caching

Vectors are written to <project>.embeddings.json (add it to .gitignore; it is 2.5 MB for 210 concepts). Only concepts whose text changed are recomputed, and a fully cached run does not load the model at all — 11s cold, 1s warm on face85.

Changing the model or the field list invalidates every vector, by design: vectors from different models are individually valid and mutually meaningless. Use --rebuild-embeddings to force a full recompute.

Querying

SELECT expand(vector.neighbors('Chain[embedding]', <query_vector>, 5))

Note: ArcadeDB stores and searches vectors but does not generate them — that is what the [embeddings] extra is for. The Neo4j backend does not support vectors in this version.

Case studies

Two real corpora, embedded and queried end to end, showing the model choice is a per-project decision rather than a fixed default.

cd face85 && synesis-graph arcadedb --project face85.synp \
    --vector-embeddings ontology_description,topic

cd Social_Acceptance && synesis-graph arcadedb --project social_acceptance.synp \
    --vector-embeddings ontology_description,topic

Querying each database with vector.neighbors:

face85, "quem manda nas decisões da empresa" (no shared word with any description):
    decisões_estratégicas       0.3288
    governança_corporativa      0.3871
    desempenho_organizacional   0.4291

Social_Acceptance, "why do people distrust offshore wind energy projects":
    Industry_Attitude           0.3451
    Preference_Misalignment     0.3529
    Offshore_Wind               0.3615

all-mpnet-base-v2 was chosen over the smaller all-MiniLM-L6-v2 for Social_Acceptance because the corpus is large and English-only, so the extra 768-dimension quality is worth the slower encode — the tradeoff table above lets you make the opposite call for a corpus where speed matters more.

Installation

Requires Python 3.10+ and synesis ≥ 0.10.0.

# Clone the repository
git clone https://github.com/synesis-lang/synesis-graph.git
cd synesis-graph

# Install (editable) with the graph backends you need
pip install -e ".[neo4j]"

# Everything except Neo4j works out of the box — including the local graph
# engine, which brings its own Java. Nothing else to install.
pip install -e .

Compatibility matrix

GDS Plugin (Optional)

For advanced metrics, install the Neo4j Graph Data Science:

# Neo4j Desktop: Plugins → Install Graph Data Science Library
# Neo4j Server: Download JAR and add to plugins/ directory

Configuration

Create a config.toml file in the root with the credentials of the backend you use:

[neo4j]
uri = "bolt://localhost:7687"  # Or your Neo4j Aura URI
user = "neo4j"
password = "your_secret_password"
database = "neo4j"             # Optional, default is 'neo4j'

[arcadedb]
uri = "http://localhost:2480"  # HTTP endpoint — NOT a bolt:// URL
user = "root"
password = "your_secret_password"
# database = "my_corpus"       # Optional, derived from the project name
# fulltext_analyzer = "brazilian"

Both blocks may coexist; each backend reads only its own. See config.toml.example for every option.

Usage

Pick a backend and point it at the Synesis project file (.synp):

# Sync to Neo4j
synesis-graph neo4j --project ./my-project/analysis.synp

# Sync to ArcadeDB
synesis-graph arcadedb --project ./my-project/analysis.synp

# Sync to a local ArcadeDB database — no server, no port, no Java
synesis-graph arcadedb-embedded --project ./my-project/analysis.synp

# Render a self-contained interactive HTML graph
synesis-graph html --project ./my-project/analysis.synp --output graph.html

# See every backend and its options
synesis-graph --help

What happens during execution?

1. Compilation: The Synesis compiler validates your code. Syntax errors are displayed and the process stops (the destination is not touched).
2. Connection: If compilation succeeds, the pipeline connects to the selected backend.
3. Constraints: Uniqueness rules are applied based on the Template.
4. Synchronization: Data is injected (Concepts, Citations, Sources, Relationships).
5. Native Metrics: Calculated via pure Cypher.
6. Advanced Metrics: GDS on Neo4j (with a warning if the plugin is absent), built-in algo.* on ArcadeDB (always available).

Asking the Graph Questions (serve)

An exported graph is a directory. serve publishes it over MCP, so Claude Desktop, Claude Code or the VSCode extension can query the corpus in natural language:

synesis-graph serve                       # publishes ./databases, Ctrl+C stops
synesis-graph serve --db-path ./graphs    # a root exported elsewhere

The command prints the mcpServers entry to paste into the chat client's configuration. Three things it handles that the engine does not:

• MCP starts disabled in the embedded distribution, and the setting lives in the running server rather than on disk — so it is enabled on every start.
• Writes are off unless --allow-writes says otherwise. A corpus is months of coding work, and reading it is the use case. Administrative operations stay blocked either way.
• The password is generated per session and printed, never written to a file. Set SYNESIS_DB_PASSWORD to keep one across restarts so the client configuration stays valid.

Serving a root with no database under databases/ is refused rather than started: a server over the wrong directory runs happily and answers every query with no rows.

Resulting Graph Structure

Main Relationships

Cypher Query Example

// Find the 10 most central concepts
MATCH (c:Concept)
WHERE c.pagerank IS NOT NULL
RETURN c.name, c.pagerank, c.mention_count, c.community
ORDER BY c.pagerank DESC
LIMIT 10

// Explore thematic communities
MATCH (c:Concept)
WHERE c.community = 42
RETURN c.name, c.pagerank
ORDER BY c.pagerank DESC

MCP Agent Integration (AI)

Once your data is in Neo4j, you can use the Neo4j MCP Server to allow LLMs (like Claude Desktop or Cursor) to converse with your research.

Quick Installation

# Install uv (if needed)
pip install uv

Claude Desktop Configuration

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "synesis-neo4j": {
      "command": "uvx",
      "args": ["mcp-neo4j-cypher@0.5.2", "--read-only"],
      "env": {
        "NEO4J_URI": "bolt://localhost:7687",
        "NEO4J_USERNAME": "neo4j",
        "NEO4J_PASSWORD": "your_password",
        "NEO4J_DATABASE": "database_name"
      }
    }
  }
}

File location:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Question Examples

Complete Documentation

See the mcp/ folder for:

• SETUP.en.md - Complete configuration guide
• QUERIES_REFERENCE.en.md - Cypher query reference
• Configuration templates for single and multiple databases

Flow Diagram

graph TD
    %% Styles
    classDef files fill:#e1f5fe,stroke:#01579b,stroke-width:2px;
    classDef engine fill:#fff3e0,stroke:#ff6f00,stroke-width:2px;
    classDef data fill:#e8f5e9,stroke:#2e7d32,stroke-width:2px;
    classDef graphDb fill:#f3e5f5,stroke:#7b1fa2,stroke-width:2px;
    classDef agent fill:#212121,stroke:#000,stroke-width:2px,color:#fff;

    subgraph "1. Input: Research as Code"
        SYN[Annotated Corpus<br/>.syn]:::files
        SYNO[Ontology<br/>.syno]:::files
        SYNP[Project<br/>.synp]:::files
        SYNT[Template<br/>.synt]:::files
    end

    subgraph "2. Synesis Engine"
        COMPILER{{synesis.load}}:::engine
        VALIDATOR(Semantic Validation):::engine

        SYNP --> COMPILER
        SYN --> COMPILER
        SYNO --> COMPILER
        SYNT --> COMPILER
        COMPILER --> VALIDATOR
    end

    subgraph "3. Structured Data"
        JSON[Canonical Object<br/>Traceable]:::data
        SCHEMA[Dynamic Schema<br/>from Template]:::data

        VALIDATOR -->|Success| JSON
        SYNT -.->|Defines| SCHEMA
    end

    subgraph "4. Knowledge Graph"
        NEO4J[(Neo4j)]:::graphDb
        NATIVE[Native Metrics<br/>Cypher]:::graphDb
        GDS[GDS Metrics<br/>Optional]:::graphDb

        JSON & SCHEMA -->|Sync| NEO4J
        NEO4J --> NATIVE
        NATIVE --> GDS

        subgraph "Metrics"
            DEG[Degree<br/>Centrality]
            PR[PageRank<br/>Relevance]
            BC[Betweenness<br/>Bridges]
            COM[Louvain<br/>Communities]
        end
        NATIVE -.-> DEG
        GDS -.-> PR & BC & COM
    end

    subgraph "5. Intelligent Consumption"
        MCP[MCP Agent]:::agent
        LLM[LLMs / Claude]:::agent

        NEO4J <-->|GraphRAG| MCP
        MCP <-->|Queries| LLM
    end

License

This program is distributed under the GNU Affero General Public License, version 3 only (AGPL-3.0-only), with the Synesis Data-Output Exception — see LICENSE and LICENSE.exception.

SPDX identifier: AGPL-3.0-only AND LicenseRef-Synesis-data-output-exception

The graphs you generate are yours. Exported artifacts — Cypher scripts, Neo4j databases, and the standalone HTML visualization — are not covered by the AGPL and carry no copyleft obligation toward Synesis.

This matters especially here: the generated HTML embeds Synesis's own JavaScript and CSS so the visualization can render standalone. That embedded material is Synesis Runtime Material under the Exception, and it does not place the generated file under the AGPL. You may publish, sell, or license the HTML you generate under any terms you choose.

The AGPL applies to synesis-graph itself: if you modify it and distribute it, or run it as a network service, you must share your changes under the AGPL.

Releases published before this change remain available under the MIT license they were issued under.

This license grants no rights to the "Synesis" name or logo.

Part of the Synesis Language ecosystem.