graph.by
We study how your firm collects, shares and uses what it knows, then build the knowledge graphs and decision systems that put it to work.
Clients, engagements, experts, filings, regulations. The answers sit in the connections between them, and most AI systems discard those connections when they split documents into chunks.
Vector search
Vector search returns six passages that sound relevant. None of them says which supplier is in scope, so the model fills the gap with a guess.
Graph traversal
Calder Batteries is in scope. Morava Inverters supplies power electronics, which the regulation does not cover.
Each relationship records the document, page and date it came from. Answers cite the passage itself, so a reviewer can check a claim instead of trusting it.
Morava Inverters was a preferred supplier in 2024 and on a watchlist by mid-2025. Ask about 2024 and you get 2024, not whichever document happens to be newest.
“Calder Batteries”, “Calder Batt. AS” and “CB Energy Storage” are the same supplier. Entity resolution merges them, so exposure, spend and history add up correctly.
Project Heron was for Halden Power, a Nordvik competitor. Teams staffed on Nordvik never retrieve it, not even through an indirect path.
Strategy consultancies, law firms, private equity, research desks and regulated industries. Anywhere a wrong answer costs more than the system.
See who we work with17-34%
of queries answered incorrectly or with misgrounded citations by leading AI legal research tools. Every tool tested was built on retrieval-augmented generation.
80% vs 51%
correct answers with GraphRAG versus vector-only RAG, across finance, healthcare, industrial and legal documents.
3 in 4
McKinsey employees use Lilli, the firm’s internal knowledge AI. Firm-wide AI is now table stakes for consultancies.
Teams whose work is reviewed, cited and billed. Each one asks questions that cross documents, dates and permissions.
“What did we recommend to utilities on grid storage since 2022, and what happened afterwards?”
Heads of knowledge, practice leaders, CIOs
A wrong figure on a slide costs a partner’s credibility in front of a client’s board.
“Which of our precedent SPAs include a reverse break fee above 3% under English law?”
Knowledge and innovation leads, practice support lawyers
A misgrounded citation is a professional-conduct problem, not a typo.
“Which of our portfolio companies share a tier-1 supplier with the target?”
Deal teams, operating partners, diligence leads
Exposure missed in diligence surfaces after signing, when it is most expensive.
“Which covered companies changed guidance within two quarters of a CFO change?”
Research directors, data and quant teams
A model that blends quarters or confuses issuers publishes the wrong call.
“Which of our product claims rely on studies superseded since the last submission?”
Regulatory affairs, compliance and quality teams
Regulators ask for the evidence trail. The graph is the evidence trail.
Every engagement starts with a Knowledge Diagnostic. We read what your firm has collected, follow how the work actually moves and watch how people use what they know. Your experts review our findings instead of filling in forms.
Weeks 1-2
We map where your knowledge lives: repositories, systems and people, how current it is and who owns it.
Weeks 2-4
We follow three to six core processes and read the questions people already ask, in search logs, tickets and help channels.
Weeks 3-5
Structured interviews with your experts about real past decisions: the cues, sources and judgement behind each one.
Weeks 5-6
Maturity scored against ISO 30401, a test set mined from real questions and checked by your experts, a baseline of your current tools, and a roadmap ranked by value.
Weeks 7-16
Ontology, ingestion, entity resolution and provenance on every edge. GraphRAG or a decision layer wired into the tools you use, and process changes where knowledge gets lost.
Weeks 17-18
We re-run the test set and the process measures, explain every miss, then hand over runbooks and governance or stay on as stewards.
Knowledge DiagnosticBuild
Before we build, we agree precision and citation thresholds with you, measured on a test set drawn from the questions your people actually ask. If the system misses them at acceptance, we fix it at our cost within a remediation period set in the contract.
Thresholds are set per engagement during the Diagnostic.
Every engagement starts with a Knowledge Diagnostic. Its fee is credited in full if you continue to a Build within 60 days.
A production knowledge graph and GraphRAG layer for one domain.
from$240,000
10 to 14 weeks
Find out what your firm knows, where it gets lost, and what is worth building.
$60,000
fixed fee, 6 weeks
A multi-domain graph with a reasoning layer for high-stakes decisions.
from$650,000
5 to 8 months
Schema changes, new sources, monthly re-scoring on your test set, freshness checks and on-call support once the graph is live.
from$18,000/ month
Prices in USD, excluding taxes, cloud infrastructure and model usage.
The paperwork follows the work. One framework agreement covers every stage, and each stage is a short order with its own scope, price and acceptance test.
Templates for every document are ready before the Diagnostic starts. We can work on your paper instead.
Usually not. We add the graph beneath it. Vector search stays useful for open-ended questions, the graph handles multi-hop, time-bound and permissioned ones, and a router decides which to use per query. Independent tests show the same split: graphs pay off on multi-step and corpus-wide questions, not on simple lookups (GraphRAG-Bench, 2026).
In your environment. We deploy into your cloud tenant or on‑premises, work under your NDA and security review, and never use client material to train models.
Whatever fits your stack. We are vendor-neutral across Neo4j, Amazon Neptune, Stardog, GraphDB and PostgreSQL-based stores, with hosted or self-hosted language models.
On a test set built from the questions your people actually ask, mined from search logs, tickets and expert interviews, then checked by your experts. We report precision, recall and citation accuracy per question type, so you can see where the system is strong and where it declines to answer.
About 10 to 15 hours each, for 8 to 15 experts across the Diagnostic: a 90-minute interview about real decisions, then a few hours reviewing the questions and answers we mined. Nobody writes a question bank from scratch.
Yes. A three-week Retrieval Audit ($35,000) scores your current system on a test set mined from your real questions and explains every failure by type. Its fee is credited toward a Build.
A first domain is typically in production 10 to 14 weeks after the Diagnostic. Further domains reuse the ontology and pipeline, so they move faster.
Yes. We work as a delivery partner when a client project needs a knowledge graph built, under your engagement lead and your client’s terms, or as a dedicated team of knowledge, ML and data engineers under your brand.
Bring one question your firm struggles to answer, or a process where knowledge gets lost. In 45 minutes we will show how a graph would handle it, and what it would take to build.