Bring your own data

You can ingest your own typed knowledge graph into OWLGraph and query it via the SDK. This page documents the “I already have typed data” path — you bring four files (ontology + entities + triples + passages); OWLGraph loads them. If you have only source material and want OWLGraph to induce the ontology + triples from it, that is the default path in the dashboard’s Ingest tab. It did not replace this one: the bring-your-own-four-files path below still exists.

What you need

File	Format	What it is
`ontology.ttl`	OWL 2 RL Turtle	Your classes + properties. Defines the typed graph schema.
`entities.jsonl`	One JSON per line	Your canonical entities + their type + aliases + the passages that mention them.
`triples.jsonl`	One JSON per line	Typed edges between entities.
`passages.jsonl`	One JSON per line	Your corpus chunks. Embeddings optional but recommended.

Total size limit today: 50 MB combined (nginx body cap). With 1536-dim embeddings, that’s room for ~3000 passages. Larger corpora need the S3-backed upload path which isn’t built yet.

File details

`ontology.ttl`

Standard OWL 2 RL Turtle. Declare your classes, properties, and (optionally) owlcore:textEmbedding if you want HNSW vector search.

@prefix : <http://your-domain.example.org/> .
@prefix owl: <http://www.w3.org/2002/07/owl#> .
@prefix rdfs: <http://www.w3.org/2000/01/rdf-schema#> .
@prefix owlcore: <https://owlgraph.ai/owl/2026/core#> .
@prefix owlv: <https://owlgraph.ai/owl/2026/vector#> .

:Drug a owl:Class .
:Disease a owl:Class .
:Approval a owl:Class .

:treats a owl:ObjectProperty ;
    rdfs:domain :Drug ;
    rdfs:range :Disease .

:approvedFor a owl:ObjectProperty ;
    rdfs:domain :Drug ;
    rdfs:range :Disease .

owlcore:textEmbedding a owl:DatatypeProperty ;
    rdfs:domain owlcore:Passage ;
    rdfs:range owlv:Float32Vector ;
    owlv:vectorDimension 1536 ;
    owlv:vectorIndex "hnsw(metric: \"cosine\", exponent: \"5\")" ;
    owlv:embedderHint "openai/text-embedding-3-small" .

`entities.jsonl`

One JSON object per line. Required: name, top_type. Optional but recommended: aliases, passage_ids.

{"name": "Pembrolizumab", "top_type": "Drug", "all_types": ["Drug"],
 "aliases": ["Keytruda", "MK-3475"],
 "passage_ids": ["doc-0001", "doc-0017", "doc-0042"]}
{"name": "Melanoma", "top_type": "Disease", "all_types": ["Disease"],
 "aliases": ["malignant melanoma"],
 "passage_ids": ["doc-0001", "doc-0042"]}

Field	Type	Notes
`name`	str	Canonical surface form. Used as the primary key.
`name_lower`	str (optional)	Lowercased. The loader fills this in if absent.
`top_type`	str	The entity’s primary ontology class. Must match a class in your `ontology.ttl`.
`all_types`	list[str] (optional)	Full type set. Defaults to `[top_type]`.
`aliases`	list[str] (optional)	Alternative surface forms for entity_search.
`passage_ids`	list[str]	Ids of passages mentioning this entity. Becomes `mentions/` edges in the graph.

`triples.jsonl`

One JSON object per line. Required: subject, subject_top_type, predicate, object, object_top_type.

{"subject": "Pembrolizumab", "subject_top_type": "Drug",
 "predicate": "treats",
 "object": "Melanoma", "object_top_type": "Disease",
 "evidence": "approved for unresectable or metastatic melanoma",
 "confidence": 1.0}

Field	Type	Notes
`subject`	str	Must match an entity name from `entities.jsonl`.
`subject_top_type`	str	The subject’s `top_type` (used to disambiguate).
`predicate`	str	An OWL ObjectProperty or DatatypeProperty from your ontology.
`object`	str	Either another entity name OR a literal value (date, number, string).
`object_top_type`	str	The object’s `top_type`. For literals, use `"Literal"`.
`evidence`	str (optional)	A source snippet supporting the triple. Surfaced in evidence chains.
`confidence`	float (optional)	0.0–1.0. Defaults to 1.0.

`passages.jsonl`

One JSON object per line.

{"id": "doc-0001", "text": "Pembrolizumab (Keytruda) is approved for...",
 "source": "FDA-Drug-Approvals/2014/pembrolizumab.pdf",
 "vec": [0.0123, -0.0456, 0.1234, ...]}

Field	Type	Notes
`id`	str	Unique chunk id. Referenced by `entities.jsonl`’s `passage_ids` and `triples.jsonl`’s `evidence`.
`text`	str	The passage content. Surfaced verbatim in `read_chunk` tool calls.
`source`	str (optional)	Document/section identifier. Shown in evidence-chain UI.
`vec`	list[float] (optional)	Pre-computed embedding. Must match your ontology’s `owlv:vectorDimension`. Omit to let the platform embed at ingest time.

Upload via the CLI

Sign up + mint an API key. Quick start covers the first half.
Install the SDK.
Terminal window
```
pip install owlgraph-core
```
Create a database.
Terminal window
```
owlgraph db create --name my-knowledge-graph --size s
```
Record the UUID it prints — you’ll pass it to subsequent commands.

Ingest your four files.

owlgraph ingest custom <db_id> \
    --ontology  ontology.ttl \
    --entities  entities.jsonl \
    --triples   triples.jsonl \
    --passages  passages.jsonl \
    --follow

--follow streams per-stage events to stdout. Without it, the command returns the job id and you tail it later with owlgraph ingest logs <db_id> <job_id> --follow.

Query it.

from owlgraph_core import sdk as owl

db = owl.connect(database_id="<db_id>", api_key="<key>")
result = db.retrieve("Which drugs treat melanoma?")
print(result.answer)
for p in result.passages:
    print(p.source, p.text[:200])

Common gotchas

invalid_jsonl in entities.jsonl line 14 — one of your JSONL rows isn’t valid JSON. Check trailing commas, unescaped quotes, missing braces.

entity name not found: <foo> — a triple references a subject or object that isn’t in entities.jsonl. Make sure every entity in the triples has a row in entities.

Schema conflicts — if the ontology declares a property with a different type than your data writes (e.g. textEmbedding as string instead of float32vector), Dgraph rejects later writes. Drop the database and re-ingest cleanly if you change the ontology.

Body too large — the API rejects total payloads >50 MB. Either drop the vec field (platform will embed at ingest cost) or split the corpus into multiple databases.

What’s not in v1

Re-ingest of just one file. v1 reloads everything per call. Re-ingest deletes existing data first (drop_all on the alpha) — back up before re-running.
S3-backed upload for >50 MB corpora. Tracked.
Induction from a raw corpus. This four-file path loads typed data you already have; it does not extract anything. If you have only source material and want OWLGraph to build the ontology and graph for you, use the dashboard’s Ingest tab, which is the default path.