Using VeritasGraph with an existing ontology or taxonomy

[astzhalkouski]

Hi! I’ve been exploring VeritasGraph and really like the direction: GraphRAG, visual graph exploration, multi-hop reasoning, citations, MCP access, and the local agent workspace all feel highly relevant for explainable AI systems.

I’m working on AI Atlas — an open, community-reviewed knowledge map of AI concepts and relationships: https://ai-atlas-project.org

One thing I’m curious about: could VeritasGraph support a workflow where the starting point is not only unstructured documents, but an already curated concept graph?

For example, AI Atlas has a manually reviewed structure of concepts and relationships. Ideally, a tool like VeritasGraph could:

• import existing nodes and edges from JSON/CSV/GraphML/Neo4j-style data,
• preserve curated relationships as authoritative,
• optionally enrich the graph with additional extracted or inferred relationships,
• support GraphRAG-style natural-language exploration,
• expose the graph through the Studio UI and MCP tools,
• and keep provenance clear: curated edge vs extracted edge vs inferred/suggested edge.

From the docs, I see the main flow is currently document/text ingestion → extracted entities/relationships → graph → query. Is there already a recommended path for importing or constraining an existing ontology/taxonomy/knowledge graph, or would that be a potential feature direction?

I think this could be useful beyond AI Atlas too — for educational knowledge maps, enterprise taxonomies, ontologies, policy graphs, and other curated knowledge bases.

[bibinprathap]

Hey @astzhalkouski! 👋

This is an excellent question, and AI Atlas looks like exactly the kind of curated knowledge graph that should plug into VeritasGraph. You've identified a really important expansion of the current flow, and I think it's absolutely the right direction.

Current State

Right now, VeritasGraph's main pipeline is indeed document → extraction → graph → query, but the underlying architecture is more flexible than the docs currently show:

• The graph backend (Parquet files for entities/relationships, optional Neo4j) can store any nodes and edges with arbitrary attributes — the extraction pipeline is just one way to populate it.
• The Studio UI and MCP tools query the graph storage directly, so they don't care whether nodes came from document extraction or manual curation.
• The GraphRAG query engine operates on the graph structure + text units, not on the extraction provenance.

So the foundation is there — what's missing is the import tooling and provenance tagging to make your workflow first-class.

What You'd Need (and what I'd add)

To make AI Atlas → VeritasGraph seamless, here's the path:

1. Import existing graphs (new module: veritasgraph.import)

Add importers for common formats:

from veritasgraph.import import import_graph

# JSON/CSV/GraphML → VeritasGraph internal format
import_graph(
    source="ai-atlas.json",
    format="json",  # or "csv", "graphml", "neo4j-dump"
    mark_as_curated=True,  # provenance flag
    merge_strategy="preserve_curated"  # curated edges win on conflict
)

[bibinprathap]

This would map your nodes/edges to VeritasGraph's entity/relationship schema and write them to the Parquet storage (or Neo4j if you're using that backend).

2. Provenance tagging (extend entity/relationship attributes)
   Add a source_type field to every node/edge:

source_type: "curated" — from AI Atlas / manual review (authoritative, never auto-deleted)
source_type: "extracted" — from document extraction (can be re-extracted)
source_type: "inferred" — from LLM suggestion or graph completion (lower confidence)
Studio UI and MCP tools would surface this (e.g., visual styling: curated edges = solid, inferred = dashed).

3. Hybrid enrichment (extend the extraction pipeline)
   Allow the extraction pipeline to augment a curated graph instead of building from scratch:

4. Studio & MCP exposure (already mostly works!)
   Once the graph is imported:

Studio UI — the graph explorer, Q&A, and playground would work immediately (they just read the Parquet/Neo4j graph).
MCP tools — veritasgraph_query, veritasgraph_get_graph, veritasgraph_search_entities already expose the full graph; no changes needed.
The only enhancement would be adding veritasgraph_import_graph and veritasgraph_export_graph MCP tools for round-trip workflows.

Timeline & Collaboration
This is not a fundamental architecture change — it's mostly new import/export utilities + provenance metadata. I'd estimate:

Import tooling (JSON/CSV/GraphML) — ~1-2 weeks
Provenance tagging (schema + UI) — ~1 week
Hybrid enrichment mode — ~2 weeks
MCP import/export tools — ~3 days
I'd love to collaborate on this. If AI Atlas is interested, here's what would help:

Sample export — a small JSON/CSV dump of AI Atlas structure (nodes: concepts, edges: relationships) so I can design the import schema.
Use-case priority — which workflow matters most first? Import-only? Import + enrich? Round-trip (VeritasGraph → AI Atlas updates)?
Provenance requirements — do you need more than curated/extracted/inferred, or is that sufficient?
I think this would be valuable way beyond AI Atlas — you're right that enterprise taxonomies, educational knowledge maps, policy graphs, and ontologies all want the same thing. Making VeritasGraph "curated-graph-first" alongside "document-first" feels like the natural evolution.

Let me know if you'd like to sync on this — happy to prioritize it if there's a concrete use case to build against. 🚀

[astzhalkouski]

Thanks, this is a very encouraging response, and the curated-graph-first direction you described is exactly what I had in mind.

For AI Atlas, I would prioritize the workflow like this:

1. Import and exploration first
   Import the existing curated graph, preserve its concept IDs and relationships, and make it available through Studio and the current MCP query tools.
2. Optional enrichment second
   Allow VeritasGraph to add extracted or inferred relationships, but keep them clearly distinguishable from the curated AI Atlas content.
3. Round-trip later, if useful
   I would not prioritize automatic updates back into AI Atlas initially. AI Atlas has a review process, so extracted or inferred additions should first be treated as proposals rather than canonical changes.

For provenance, curated, extracted, and inferred are a good starting point. The main requirement on our side is that curated relationships remain authoritative and that generated additions never become indistinguishable from reviewed AI Atlas content. Preserving source information such as the original concept ID and source version would also be valuable.

The current canonical AI Atlas JSON is here:

https://github.com/ai-atlas-project/ai-atlas/blob/main/taxonomy/ai-taxonomy-l1-l2.json

I can also prepare a smaller sample containing one branch of the taxonomy, together with the expected nodes, hierarchy relationships, and metadata. That may be easier for designing and testing the first importer.

A focused first milestone could simply be:

Import a small AI Atlas JSON graph, preserve its curated structure and metadata, and query it through Studio and MCP.

That would already demonstrate the main use case without requiring enrichment or round-trip updates yet.

I’d be happy to prepare the sample and test an early implementation. Thanks again for being open to collaborating on this.

[bibinprathap]

Hey @astzhalkouski! 👋

Great — this aligns perfectly with where we're taking Veritas Studio.

Concrete plan: Veritas Studio will have a built-in option to upload a pre-built knowledge graph directly, rather than only building one from document ingestion. That means you'll be able to bring:

Knowledge graph output from Graphify
Graph output from Understand-Anything
Or any other JSON graph output — including the AI Atlas JSON graph
Once uploaded, the imported graph becomes a first-class citizen in the orchestration framework — it feeds straight into the GraphRAG query engine, the Studio graph explorer, and the existing MCP tools (veritasgraph_query, veritasgraph_get_graph, veritasgraph_search_entities), with no changes needed on the query side.

This maps directly onto the milestone you outlined:

Import & exploration first — upload the curated AI Atlas JSON, preserve original concept IDs, hierarchy, and metadata, and immediately explore/query it through Studio + MCP. ✅
Optional enrichment second — extracted/inferred relationships get added on top, tagged distinctly (curated / extracted / inferred) so reviewed AI Atlas content never becomes indistinguishable from generated additions. Curated edges stay authoritative.
Round-trip later — treated as proposals only, deferred as you suggested.
On provenance: curated / extracted / inferred will be the baseline, and we'll preserve source info (original concept ID + source version) on every node/edge so the AI Atlas origin is always traceable.

A smaller single-branch sample (nodes + hierarchy relationships + metadata) would be ideal for designing and testing the first importer against — if you can prepare that, I'll wire up the upload path in Studio so we can validate the "import → explore via Studio & MCP" milestone end-to-end.

Excited to build this against a real curated graph. 🚀

[bibinprathap]

Hey @astzhalkouski! Quick update—I’ve officially implemented this feature. 🚀

VeritasGraph Studio now fully supports importing pre-built knowledge graphs directly from formats like Graphify, Understand-Anything, and your AI Atlas taxonomies. The core implementation covers the exact milestones we discussed:

Provenance-Aware Import: Every imported node and edge is automatically tagged as Curated, Extracted, or Inferred. The Graph Explorer visually renders these tiers cleanly, showing solid green for curated facts, dashed for extracted data, and dotted for inferred guesses.

Merge Conflict Protection: The "Preserve curated on conflict" strategy is completely operational to ensure your authoritative AI Atlas structure is never silently overwritten by lower-trust data.

Cited Reasoning: You can immediately jump into the Playground to run GraphRAG queries over your imported data and get fully grounded answers backed by traceable citations.

I’ve recorded a quick video walkthrough demonstrating the exact workflow from the initial file upload to running cited reasoning queries. You can check out the video here: Json to AI Data

Please give it a spin with the AI Atlas JSON taxonomy and let me know how it works for you!