Point us at your sources.
PDFs, docs, wikis, transcripts — upload the text your agent needs to know. No schema design, no chunking strategy, no query language to learn.
AI infrastructure · v1.0 generally available Point it at your sources.
Point it at your sources.
We do the rest.
OWLGraph turns your documents into a knowledge service any LLM or agent can plug into. Upload your sources; we build the graph, the reasoning, and the endpoint — and every answer your AI gives cites where it came from.
How it works
Three steps. No new stack.
You bring the documents. OWLGraph builds the knowledge service and hands you the endpoint.
We build the brain.
OWLGraph reads your corpus and builds a typed knowledge graph from it — the entities, the relationships, and the reasoning to traverse them. About 90 seconds to a working graph.
Point your agent at the endpoint.
You get an MCP endpoint any LLM or agent can use — Claude, GPT, or your own. One config block, and your AI answers from your knowledge and cites its sources.
The method · OwlGraphRAG Retrieval that shows
Retrieval that shows
its work.
We call the method OwlGraphRAG. Instead of guessing across text chunks, it walks a typed knowledge graph built from your sources — and returns every edge it walked. Multi-hop questions stop being a guess; they become a chain you can read, audit, and replay in your evals.
Entities resolved by type, not keyword
Edges chosen by the schema, not by similarity
Source passage cited per hop
On a controlled 100-question benchmark, OwlGraphRAG scored +11.3pp over naive vector RAG — same data, same reader model, only the retrieval differed. Read the full experiment →
Retrieval trace Trace · example
Question
Which suppliers ship to states where we're licensed?
01 Acme Corp → ships_to → California q3.pdf:14
02 California ⊆ LicensedStates inferred
03 Beta Inc → ships_to → Texas q3.pdf:22
04 Texas ⊆ LicensedStates inferred
Answer Acme Corp, Beta Inc
OWLGraph · 2 hops · 4 nodes Vector RAG returned 4 unrelated chunks
Run a question like this in the live demo → Where flat retrieval breaks
The questions your users actually ask.
Multi-hop and constraint-based queries are where evals plateau. They're also where customers notice when the answer is wrong.
Which suppliers ship to states where we're licensed?
Why it's hard Two-hop: supplier → ships_to → state, then state ⊆ licensed_states.
How vector RAG fails Vector RAG returns four chunks about shipping or licensing — never the join.
Show me drug pairs approved for X but contraindicated with Y.
Why it's hard Constraint join across two schemas — therapeutic class + interaction profile.
How vector RAG fails Embedding similarity can't enforce "approved for X AND not in interaction set."
Which of our customers have been with us 5+ years and use product Z?
Why it's hard Temporal predicate + product attribute, both must hold.
How vector RAG fails Chunked text doesn't survive constraint composition; agent loops out before converging.
In your code
from owlgraph_core import sdk as owl
db = owl.connect(
database_id="your-database-uuid",
api_key="sk-owl-...", # or set OWLGRAPH_API_KEY
)
result = db.retrieve("Which suppliers ship to states where we're licensed?")
print(result.answer)
for p in result.passages:
print(p.source, p.text[:200]) Two ways in
One engine. A depth dial,
not two products.
Start with the managed path — most teams never need more. When you want your hands on the machinery, the deeper controls are already there, on the same graph and the same endpoint.
Managed · do it for me
We do the rest.
Upload your sources and point your agent at the endpoint. You build the agent; we run its knowledge.
- Sources in, typed knowledge graph out — about 90 seconds
- Cited answers out of the box, on every question
- MCP endpoint for Claude, GPT, or your own agent
Control · open the hood
Everything above, plus the knobs.
The same engine exposes its deeper layers when you want them — no separate product, nothing to migrate.
- Inspect and edit the graph and the schema behind it
- Tune how retrieval traverses your data
- Same graph, same endpoint, at any depth
This is how FiveLoaves — our own sermon-prep product — runs in production. We're customer zero, and the control tier exists because we needed it ourselves.
XS Instance Free 2 GB RAM · 1 CPU
First instance free.
Compute from $0.25/hr per instance
Scale to zero anytime.
Storage $0.10/GB/mo
Pay for what you store.
Point it at
your sources.
Start free — no credit card. About 90 seconds from upload to a knowledge service your agent can query, with every answer cited.