October 10, 2026 · GRAFOLIO
Portfolio Graph, Not Folder Search
Portfolio Graph, Not Folder Search
RAG over a data room finds documents. A portfolio needs entities and relationships you can traverse and cite.
Pramod Thaz · Grafolio
"We put the data room in a vector store" is a common AI pilot. It is also why many portfolio questions still end in a pile of loosely related PDF snippets.
Retrieval-augmented generation — search over chunks of text, then ask a model to answer from what came back — works well for narrative research. It is a weak foundation for institutional portfolio management. A CIO or head of asset management does not only want documents that mention Beacon Wharf. They want the fund that owns it, the units and leases inside it, the statement lines that make up its NOI, and a path from each reported figure back to the source.
That is a graph problem, not a folder-search problem.
What folder search gets you
Drop rent rolls, T-12s, offering memos and investor packs into a shared drive. Index them. Ask: "What is NOI for the portfolio?"
A good retrieval system will surface pages that mention "NOI" and property names. A fluent model will assemble a paragraph. What you often cannot get reliably is whether every property was included; whether each NOI matches the firm's approved definition; whether periods align; whether a missing line was estimated or flagged; and which exact line on which statement produced each number.
Folder search finds text that looks relevant. Portfolio work needs records that are related.
What a portfolio graph is — in plain English
A knowledge graph, for our purposes, is a connected map of the things your portfolio is made of, and how those things relate.
Think of each important object as a record — a node — and each relationship as a link:
- A fund owns properties
- A property has units
- A unit has a lease (and a tenant)
- A property has operating statements
- An operating statement is made of line items (NOI, GPR, expenses, and so on)
- Documents and workbook outputs attach to those records as sources
When you ask a question, the system does not only search for similar paragraphs. It follows the links: fund to properties, property to statement lines, line to the citation on the page. The answer can show the path it took — which is what a reviewer means by "show your work."
That is the whole idea. Entities, relationships, and a trail you can audit.
Why relationships change the answer
Take an illustrative demo portfolio — call it Alder — with three multifamily assets: Beacon Wharf, Juniper Court and Vesper Lofts.
Folder search might return three NOI mentions from three PDFs and sum whatever numbers appear nearby. If Vesper Lofts has no NOI line, a chatty system may still "helpfully" fill a gap from a deck or a prior period.
Graph traversal starts at the fund, walks to each property, looks for the approved NOI line on each operating statement, and reports:
- Beacon Wharf: $1,310,000 — cited to the statement
- Juniper Court: $764,000 — cited (different trailing period, still visible)
- Vesper Lofts: not reported — line absent; flagged, not estimated
Same question. Different control outcome. The graph makes the missing relationship visible. Folder search often papers over it with fluent prose.
Questions that need a path, not a paragraph
These are ordinary desk questions. They are awkward for pure retrieval and natural for a connected portfolio record:
Acquisitions. "Which leases expire in the next ninety days, and what is contract rent?" You need property → units → leases → dates and rent, each cited to the rent roll — not a paragraph that vaguely mentions expirations.
Asset management. "Where did budget-versus-actual variance concentrate?" You need properties linked to statement lines and the prior budget, with exceptions held when totals do not tie.
Fund and LP reporting. "Roll property-level NOI to the fund." You need ownership links, consistent metric definitions, and a trail from the fund figure back to each source line.
Credit. "Which loans have covenant tests due, and which borrower packs are late?" You need borrower → facility → covenant → reporting obligation, not a keyword hit on "covenant."
In each case the value is the traversal: move along known relationships and stop when a link or field is missing.
RAG is still useful — as a layer, not the ledger
None of this argues against retrieval. Offering memoranda, IC memos and market decks are narrative; searching them by meaning is right for synthesis and diligence notes.
The mistake is treating the data room index as the portfolio's system of record. Narrative search and structured portfolio records solve different jobs:
| Job | Folder / document search | Portfolio graph |
|---|---|---|
| Find related narrative | Strong | Supporting |
| Answer "what does this memo say?" | Strong | Supporting |
| Answer "what is fund NOI by property?" | Weak | Strong |
| Enforce firm metric definitions | Weak | Strong |
| Cite a specific statement line or lease | Fragile | Native |
| Flag missing entities or lines | Easy to miss | Explicit |
| Survive IC / audit / exam review | Hard | Designed for it |
Use retrieval for documents that should stay documents. Use a connected record for the entities your reports, models and agents will treat as facts.
Definitions sit on the graph, not in the prompt
Even a perfect map of funds and leases fails if "occupancy" means three different things by strategy. Firm definitions — what counts in NOI, how vacant units are rented for underwriting, how renewal trade-out is measured — belong with the portfolio record as approved rules, not as wording buried in a prompt that changes by session.
When a question asks for a metric, look up the approved definition, follow it, and refuse to invent when the line is absent. That discipline is easier when metrics attach to named entities (this property, this period, this statement line) than when every answer is assembled from whichever chunks search returned today.
What "cited" means when the portfolio is connected
Citation is not a footnote aesthetic. It is the ability to point from a reported number to the entity (property, lease), the period, the source artifact (statement, rent roll page, workbook output), and optionally the path (fund → property → statement line).
That is what investment committees and compliance teams mean by a source trail. A vector hit that "supports" an answer is not the same thing. Supporting quotes help; structured provenance decides whether the number can leave the pilot.
From folders to a portfolio you can ask
If your AI roadmap stops at "chat with the data room," you get faster literature review and the same manual roll-ups you have today.
Invest in a portfolio graph — entities for funds, properties, units, leases and statement lines; relationships that mirror how the portfolio works; citations that travel with every figure — and agents have something solid to stand on. They can traverse, apply approved definitions, drive the Excel models that remain your system of record, and stop cold when the graph has a hole.
That is the difference between searching a filing cabinet and knowing the portfolio.
Grafolio builds agent workflows on that principle: your portfolio as a connected graph of entities, documents, models and decisions. Meridian by Grafolio is our AI-first CRE portfolio platform and reference experience — it organizes rent rolls, operating statements and deal documents into a knowledge graph of the portfolio so questions can traverse funds, properties, units, leases and statement lines, with every figure cited to its source.
Technology and process advisory. Not investment, legal or regulatory advice. Illustrative demo examples only; no client results are implied.