Why a knowledge graph

Why reasoning over a knowledge graph of the record is sturdier, cheaper and deeper than having an AI model reread the documents for every question.

There are two ways to put intelligence on top of thousands of regulatory and institutional documents. You can hand the documents to an AI model every time someone has a question. Or you can read them once, carefully, into a knowledge graph, and reason over that. EphemerisX does the second. This page explains why, and shows what it finds that the first approach misses.

Two approaches#

Figure 1. Re-reading the documents for every question, and reasoning over a graph built from them once.

Why the graph is more robust#

Every fact has an address. Each entity and relation in the graph points to the document and page it came from. An answer built on the graph can cite every step, and anyone can check it. See Verify a quote.
The same question gets the same answer. The graph does not change between questions unless the record does. A model re-reading documents can weigh them differently each time.
Gaps are visible, not filled. When the record is silent, the graph has no edge, and the sites say so (“[no position recorded on this thread]”, Record incomplete). A model asked the same question tends to fill the silence.
Changes are tracked. When FDA posts a new letter or a company files, the new document is added to the graph and its effects are shown. Nothing already read has to be read again.

Why it is more efficient#

The expensive part of working with documents is reading them. A graph pays that cost once; a model that re-reads pays it on every question.

On the FDA Ephemeris at the time of writing, 13 documents (939 pages) had been read into 1,516 entities and 2,346 relations, out of 130 documents (5,954 pages) in the system. A question across everything held would mean re-reading roughly 6,000 pages, on the order of three million words, for each question. Answered from the graph, the same question touches only the entities and relations that matter to it, each already linked to its page.

	Re-reading with a model	Reasoning over the graph
Reading cost	Paid on every question	Paid once per document
Growth	Grows with pages × questions	Grows with new documents only
Scope of one answer	What fits in one reading	The whole record, every domain
Where effort goes	Finding the relevant passages again	Comparing, connecting and testing them

Why it goes deeper#

Because the reading is already done, the reasoning budget goes where the value is:

Connections: The same entity links documents from different domains. An airport code in a flight record, a counterparty in an SEC filing and a name in a shareholder letter meet at one node. See the private-jet example.
Relationships: Each decision's positions are compared with earlier ones on the same question: consistent, conditional, or departing. See the threads.
Differences: Small changes in an agency's wording across years of documents are found by comparing linked positions, not by rereading every label.
Weaknesses: Where a decision rests on a single piece of evidence, where a committee dissented, or where the record has no position at all, the graph shows it as structure: a thin chain, a split vote, a missing edge.

Example: how FDA's ApoE ε4 guidance hardened#

Carriers of the ApoE ε4 gene have a higher risk of ARIA, the brain swelling and bleeding seen with anti-amyloid antibodies. Across three decisions and two and a half years, FDA's labels moved from noting that risk to requiring a test for it. The graph finds this by following one thread (T3, ARIA) and comparing each position with the one before.

Jun 7, 2021 · Aduhelm label

Origin

“The incidence of ARIA-E was higher in apolipoprotein E ε4 (ApoE ε4) carriers than in ApoE ε4 non-carriers (42% and 20%, respectively).”

Label · PDF p. 5 · position ADU-T3-1

departs

Jan 6, 2023 · Leqembi label

Departs

“Consider testing for ApoE ε4 status to inform the risk of developing ARIA when deciding to initiate treatment with LEQEMBI.”

Label · PDF p. 5 · position LEQ-T3-3

departs

Jul 6, 2023 · Leqembi label, traditional approval

Departs

“Testing for ApoE ε4 status should be performed prior to initiation of treatment to inform the risk of developing ARIA.”

Label · PDF p. 2 · position LQT-T3-2

Figure 2. One thread, three linked positions: from noting the risk, to “consider testing”, to “should be performed”. Each position is cited to its FDA label page.

Along the same thread, the graph also shows what did not change: MRI monitoring before early infusions carries through from Aduhelm to Leqembi (Consistent), and Kisunla's and Leqembi Iqlik's labels keep the heightened risk for ApoE ε4 homozygotes.

Why this is hard without a graph#

The three sentences sit in three different labels, published years apart, among more than 30 documents behind these decisions.
They do not share wording a keyword search would match: “incidence was higher”, “consider testing”, “should be performed”.
The change only shows when each position is compared with the one before it on the same question, which is exactly what the graph's links record.

For an investor or a sponsor, that shift is a signal about the next label in the class: testing requirements change who can be treated, how quickly, and at what cost.

See the full chain on the threads page guide, or open it live at alzheimers.ephemerisx.com/threads/#T3.

Example: a weakness in the record#

Aduhelm's accelerated approval rested on reduction of amyloid on PET imaging as a surrogate (thread T1, Origin). Its advisory committee had voted 0 yes, 10 no and 1 uncertain. In the graph, that decision is the one place in the anti-amyloid record where an approval node links to a committee vote against it, and where the surrogate thread starts. That structure is what a risk review looks for: a decision whose support rests on one contested line of evidence. See all four gaps in Aduhelm’s approval. See Committee vs. approval and Risk analysis.