Output Formats

Output Formats

Synaptic builds one in-memory knowledge graph and can serialize it to many artifacts. The canonical artifact is graph.json (a NetworkX node-link document); everything else is derived from the same graph.

Two ways to produce artifacts:

• synaptic extract writes a full default set in one pass (see [Extraction]).
• synaptic export <format> regenerates a single artifact from an existing graph.json without re-extracting, or pushes the graph live to a database.

For the visual/interactive artifacts (graph.html, graph-3d.html, graph.svg, callflow.html, tree.html) see [Visualizations].

What extract writes by default

synaptic extract writes all of these into synaptic-out/:

graph.json        graph.html        GRAPH_REPORT.md
graph.graphml     graph.cypher      graph.dot
callflow.html     tree.html         graph.svg
graph-3d.html

--obsidian additionally writes synaptic-out/obsidian/ and --wiki writes synaptic-out/wiki/. Neither is produced by default.

Bulk exports are skipped on very large graphs

graph.graphml, graph.cypher, graph.dot and graph-3d.html emit every node and every edge with no aggregation, so their size tracks the graph exactly. Past 150,000 nodes they are skipped and extract says so, because at that scale they are large enough that no tool in the chain will open them (a 379k-node graph produced a 342 MB GraphML and a 259 MB graph-3d.html).

graph.json and graph.html are always written: graph.json is the canonical contract, and graph.html aggregates by community above 5,000 nodes so it stays responsive.

The export command

synaptic export <format> [--graph <path>] [--out <path>] [--repo <tag>]
                          [--push <uri>] [--user <u>] [--password <p>]

export loads an existing graph.json, then re-emits one format from it. No re-extraction happens, so it is fast and deterministic.

Flags:

• --graph <path> source graph. Default: synaptic-out/graph.json.
• --out <path> output file (single-file formats) or directory (obsidian/wiki). Default: alongside the source graph, using the conventional filename for that format.
• --repo <tag> scope to one federated member by its repo tag before exporting (see [Workspaces-and-Federation]).
• --push <uri> for neo4j/falkordb only: push live to a running database instead of writing the cypher script. Requires the push build feature.
• --user <u> auth user for --push to Neo4j (default neo4j).
• --password <p> auth password for --push (or set NEO4J_PASSWORD / FALKORDB_PASSWORD).

Full format list

Verified from the command dispatch, export accepts exactly these formats (case-insensitive):

Format	Output	Notes
json	graph.json	Pretty-printed node-link JSON
html	graph.html	Interactive 2D explorer (see [Visualizations])
svg	graph.svg	Static layout (see [Visualizations])
graphml	graph.graphml	Gephi / yEd / GraphML tools
cypher	graph.cypher	Neo4j / FalkorDB import script
dot	graph.dot	Graphviz DOT
callflow (alias callflow-html)	callflow.html	Mermaid call-flow (see [Visualizations])
tree	tree.html	D3 file tree (see [Visualizations])
3d (alias force3d)	graph-3d.html	3D force graph (see [Visualizations])
obsidian	obsidian/ directory	Obsidian vault
wiki	wiki/ directory	Markdown wiki
report	GRAPH_REPORT.md	Analysis report (see [Analysis-and-Reports])
neo4j	graph.cypher or live push	See live push below
falkordb	graph.cypher or live push	See live push below

Any other value is rejected with an error listing the valid formats.

--repo scoping guard

export json --repo X with no --out is refused, because its default output name (graph.json) would overwrite the source graph with a scoped subgraph. Pass --out <path> to write the scoped graph elsewhere. Other formats have distinct default names and are unaffected.

graph.json (the canonical artifact)

graph.json follows the NetworkX node_link_data(G, edges="links") shape. Top-level keys:

{
  "directed": false,
  "multigraph": false,
  "graph": {},
  "nodes": [ ... ],
  "links": [ ... ],
  "hyperedges": [],
  "built_at_commit": "deadbeef"
}

• Edges are under links (not edges). On input, an edges key is accepted as an alias.
• hyperedges is always present, even when empty.
• built_at_commit is omitted when unknown.

Node shape

Required keys are always present; optional keys are omitted when unset. Any other keys round-trip through a flattened extra map (so re-importing is lossless).

{
  "id": "auth",
  "label": "auth.py",
  "file_type": "code",
  "source_file": "src/auth.py",
  "source_location": "L42",
  "community": 3,
  "repo": "app",
  "norm_label": "auth.py"
}

• id stable node identifier.
• label display name.
• file_type one of code, document, paper, image, rationale, concept (lowercase).
• source_file originating file path.
• source_location optional (e.g. a line marker); omitted when unset.
• community optional cluster id (integer); omitted when unset.
• repo optional federation tag; absent for single-repo graphs.
• kind optional node kind (class, function, method, struct, ...); set for code nodes in supported languages, omitted otherwise. See [Extraction].
• visibility optional declared visibility (public, private, protected, internal); omitted when unknown.
• span optional source range object (start_line, start_col, end_line, end_col); omitted when unknown. Lines-of-code is end_line - start_line + 1.
• norm_label lowercased search key, added by the JSON writer on export.

The kind (the node's NodeKind: table/column/view/function/class/… ) drives node shape and color-by-kind in the SVG, 2D, and 3D viewers, so the SQL and cross-language layers are visible — not just "code". SQL nodes also carry the facts the viewers surface on hover (dialect, data_type, pk, fk_target, rls_enabled, security_invoker). Non-code asset nodes carry asset_kind (e.g. stylesheet, data, image, font, media) for their own shapes; federated external-package stubs carry external_package: true.

Link (edge) shape

{
  "source": "a",
  "target": "b",
  "relation": "calls",
  "confidence": "EXTRACTED",
  "source_file": "src/a.py",
  "source_location": "L10",
  "confidence_score": 1.0,
  "weight": 1.0,
  "context": "...",
  "sites": [
    { "source_file": "src/other.py", "source_location": "L22" }
  ],
  "cross_repo": true
}

• source / target endpoint node ids.
• relation relationship name (e.g. calls, contains, method, imports).
• confidence one of EXTRACTED, INFERRED, AMBIGUOUS (uppercase).
• source_file where the relationship was found.
• confidence_score numeric score. The JSON writer defaults it from the confidence level when unset: EXTRACTED -> 1.0, INFERRED -> 0.5, AMBIGUOUS -> 0.2.
• weight defaults to 1.0.
• source_location, context optional, omitted when unset.
• sites optional: additional distinct extraction locations for the same semantic edge. The primary location remains in source_file / source_location; federation repo-prefixes both the primary and every additional site. Context is part of semantic identity, so otherwise-equal context-bearing edges are not merged.
• cross_repo omitted when false; true marks a federated cross-repo edge.

Hyperedge shape

{ "id": "he1", "label": "auth cluster", "nodes": ["a", "b"] }

Optional relation and confidence keys are omitted when unset.

graph.graphml

GraphML for Gephi, yEd, or any GraphML consumer. The document declares <key>s and emits <node>/<edge> data. Attributes preserved:

• Node: label, file_type, source_file, community (declared as long; -1 when unset).
• Edge: relation, confidence.

For federated graphs only, two extra attributes are declared and emitted: a node repo (string) and an edge cross_repo (boolean, written only on cross-repo edges). Single-repo GraphML is unchanged. When any node carries enrichment metadata, three node attributes are also declared and emitted: kind, visibility, and loc (long). Cypher's rich (live-push) form sets the same n.kind/n.visibility/n.loc. Internal _-prefixed extras are dropped. All text is XML-escaped.

graph.cypher

A Cypher import script for Neo4j or FalkorDB, generated as ;-terminated statements: node MERGEs first, then edge MERGEs.

Both forms are idempotent on re-run. The MERGE key is the id alone, and label/confidence are applied with SET, so re-importing after a label or confidence change updates in place instead of duplicating. Node file type becomes the node label (capitalized, allowlisted; falls back to Entity), and the relation becomes the relationship type (uppercased, allowlisted; falls back to RELATES_TO).

MERGE (n:Code {id: 'a'}) SET n.label = 'A';
MATCH (a {id: 'a'}), (b {id: 'b'}) MERGE (a)-[r:CALLS]->(b) SET r.confidence = 'EXTRACTED';

String values are escaped for single-quoted Cypher literals (drops control bytes; escapes \, ', newlines, CRs). Identifier-position tokens (node label, relationship type) are allowlisted to [A-Za-z0-9_] with a required leading letter, since Cypher cannot escape those positions.

The static script carries id + label on nodes only. The live push (see below) uses a richer form that also writes source_file and community onto nodes.

graph.dot

Graphviz DOT. Renders a digraph (with ->) when the graph is directed, otherwise a graph (with --). Layout hints rankdir=LR and node [shape=box] are emitted. Each node carries label and kind (the file type); each edge carries label (the relation).

graph Synaptic {
  rankdir=LR;
  node [shape=box];
  "a" [label="A", kind="code"];
  "a" -- "b" [label="calls"];
}

Node ids and all attribute values are emitted as double-quoted DOT strings with ", \, and newlines escaped, so an arbitrary label cannot break the statement.

Obsidian vault (--obsidian / export obsidian)

to_obsidian writes a directory (default synaptic-out/obsidian/) containing:

• One Markdown note per node, with YAML frontmatter (id, file_type, community, source_file) and neighbors as [[wikilinks]] grouped by relation. Incoming edges are grouped under <relation> (in).
• One _Community-<id>.md overview note per community, with cohesion, a member list, a Dataview live query (LIST FROM "" WHERE community = N), and links to related communities by shared-edge count.
• .obsidian/graph.json, which colors the Obsidian graph view by community via a per-note frontmatter search.

Note filenames are sanitized from labels and deduped (colliding labels get a (N) suffix), and wikilinks target the actual deduped filename so they resolve. community_labels supply semantic community names; when empty, names fall back to Community N. The function returns the count of notes written.

Markdown wiki (--wiki / export wiki)

to_wiki writes a directory (default synaptic-out/wiki/) of GitHub-wiki-style Markdown with [[wikilinks]]:

• index.md listing communities and the top "key nodes" (highest degree).
• One community-<id>.md article per community, with cohesion score, key concepts (highest-degree members), spanned source files, related communities by shared-edge count, and an audit trail (the EXTRACTED/INFERRED/AMBIGUOUS mix of the community's internal edges).
• One page per "god node" (the top 10 highest-degree nodes overall), grouping neighbors by relation.

God-node filenames are deduped the same way as Obsidian notes. The function returns the count of files written.

GRAPH_REPORT.md (export report)

A Markdown analysis report. export report recomputes communities and analysis from the loaded graph (these are not stored in graph.json) before writing. See [Analysis-and-Reports] for its contents.

Neo4j / FalkorDB live push (push feature)

export neo4j and export falkordb write graph.cypher by default. With --push <uri> and a Synaptic binary built with --features push, they stream the graph into a running database instead. Both use the richer (props + community) idempotent MERGE statements, so re-runs upsert and never duplicate.

synaptic export neo4j    --push bolt://localhost:7687 --password <pw>
synaptic export falkordb --push falkordb://localhost:6379

Transports:

• Neo4j uses cypher-shell (must be on PATH; speaks Bolt). The ;-terminated script is piped to its stdin. The URI is a Bolt URL (e.g. bolt://localhost:7687). Auth: --user (default neo4j) and --password or NEO4J_PASSWORD.
• FalkorDB uses the pure-Rust redis client, running each MERGE through GRAPH.QUERY synaptic <stmt>. The URI accepts falkordb://, redis://, rediss://, a bare host:port, or a bare host (default port 6379). Auth password is optional, taken from --password or FALKORDB_PASSWORD, and is set as a connection field (never interpolated into a URL), so passwords containing @/:///% connect correctly.

On success each prints the number of statements pushed. Without the push feature, --push fails with a clear message. See [Ingestion] and [MCP-Server] for related integrations.

See also

• Building on Synaptic for consuming graph.json as an API from any language (and building your own MCP server on it).
• [Visualizations] for the interactive and static visual outputs.
• [Analysis-and-Reports] for GRAPH_REPORT.md.
• [Commands] for the full CLI reference.
• [Workspaces-and-Federation] for --repo scoping and cross-repo graphs.