Graph RAG versus vector RAG for business files.

Business files are not a blog corpus. A matter folder holds people, documents, deadlines, sites, orders, and the links between them. When you ask a model a question over that pile, the retrieval shape decides whether the answer points at a real object — or at a fluent paragraph that could belong to any file.

Retrieval-augmented generation (RAG) means: pull relevant pieces from your own store, then let the model write language over those pieces. The arithmetic of ranking and join stays in code. The model does not invent the figure. What changes is how you pull.

Vector RAG for business files

Vector RAG embeds chunks of text and retrieves by similarity. It fits when the question is close to one passage: a clause in a contract, a paragraph in a handbook, a line in a policy, a note that already states the fact you need.

For business files that is the common path. You chunk PDFs and mail exports. You store vectors. You ask. You get the nearest chunks and citations. When the answer lives inside one document, vector retrieval is usually enough — and it is the right first build.

Limits show up when the question spans objects. “Which open deadline sits on the same matter as this supplier, and who last approved the change?” is not a single-chunk lookup. Similarity can surface related text. It does not reliably walk the relationship you already know exists in the operation.

Graph RAG for business files

Graph RAG retrieves over named entities and relationships — a case graph, an object graph of the matter or the floor. People, documents, deadlines, sites, orders, approvals sit as nodes and edges. The model still writes language. The graph supplies the path the question must follow.

That shape fits multi-hop questions over matter files and operating files: which document belongs to which matter, which person owns which approval, which site shares the same constraint. See sensitive data needs a case graph and objects from the operation — cite those cousins; this piece is the retrieval compare, not a remake.

Graph RAG is not a better chat skin. Without honest objects, a graph is decoration. With objects, a proposal can point at something you can release or refuse.

When each shape fits

Use vector RAG when:

• The answer is likely inside one document or a small set of similar passages.
• You need citations from prose you already hold.
• You are starting: files in, chunks out, no mature object model yet.

Use graph RAG when:

• The question names relationships (matter ↔ person ↔ document ↔ deadline ↔ site).
• One hop is not enough; the next fact depends on the previous object.
• Approval and access must follow the same edges as the files (who may see which node).

Many production setups run hybrid: vector search for candidate passages, then graph expansion to the entities those passages belong to — or the reverse, graph-first to scope the matter, then vector inside that scope. The product question is not “which brand of index.” It is which question shape your files actually answer.

Approval still holds the switch

Retrieval does not replace release. Agents on Spectacle may propose language grounded in retrieved objects. Without approval the agent sleeps — without approval the agent sleeps. Sealed systems of record are not written back. Models hosted on sovereign infra such as STACKIT can be a residency path for sensitive corpora; hosting is not the operation — models on STACKIT are not the operation. Cite only; do not republish that cousin.

Arithmetic stays in code. The model only writes language. We do not invent hours, ROI, euros, or accuracy bands.

What we are not selling

Not a slogan that “graph is not vector.” Not a remake of the live case-graph or STACKIT cousins. Not restaurant software or “AI for hospitality.” Not a deep-tech kernel pitch. Not a named competitor stack. Not silent write-back. Not legal advice. Counsel owns secrecy and professional-law pathways; day-2 on the Maggy calendar covers attorney service-provider rules when HoM opens it.

One core, every kind of business — one core, every operating type. Six sector bands equal.

FAQ

What is vector RAG for business files?
Similarity retrieval over embedded chunks of your documents, then language over those chunks with citations. Strong for single-passage facts.

What is graph RAG for business files?
Retrieval over entities and relationships in a case / object graph so multi-hop questions can follow real links between people, documents, deadlines, and sites.

Which should we start with?
Start with vector RAG if your files are prose and the questions are single-hop. Add a graph when questions regularly cross objects you already name in the operation.

Is hybrid required?
Often useful: vector candidates plus graph scope, or graph scope plus vector inside the matter. Not a volume claim — a shape choice.

Do agents answer without approval?
No. They propose. You release. Without approval the agent sleeps.

Is this only for law firms?
No. Matter files and operating files share the pattern wherever objects and relations matter. Six bands equal. Law is one configuration, not the only reader.

Where should I start?
Spectacle. The Insights series is how we write about the operation.

CTA

If your questions over business files still fail because retrieval cannot walk from person to document to deadline, request an initial analysis. Twenty minutes is enough to see whether vector, graph, or hybrid fits the files you already hold. If the data cannot answer, we say so.