How Evidoc Finds Answers Across Documents: Knowledge Graphs and Mathematical Reasoning

Beyond keyword search, beyond vector similarity

When you ask a question about your documents, most AI tools do one of two things:

Keyword search — find documents containing the words in your question
Vector similarity — find documents whose meaning is close to your question

Both approaches find relevant documents. Neither finds the specific sentence that answers your question, and neither follows multi-step reasoning chains across documents.

Evidoc does both. Here’s how.

Step 1: Building the Knowledge Graph

When you upload documents, Evidoc doesn’t just store them — it reads every sentence and extracts the entities and relationships it contains.

A sentence like “Acme Corporation shall pay a consulting fee of $150 per hour” produces:

Entities: Acme Corporation, consulting fee, $150 per hour
Relationships: Acme Corporation → pays → consulting fee; consulting fee → amounts to → $150/hour

These entities and relationships become nodes and edges in a Knowledge Graph. When the same entity appears in multiple documents — “Acme Corp” in the contract, “ACME Corporation” on the invoice — the graph connects them automatically.

The result: a web of connections across all your documents, built without any manual setup.

Step 2: Fine-grained indexing

Unlike document-level retrieval, Evidoc indexes your content at a fine-grained level. Each passage becomes a node in the graph, connected to the entities it mentions.

This means when you ask “What is the hourly rate?”, we don’t return a 40-page contract. We return the specific passage on page 7 that states the rate — and the passage in the amendment on page 3 that changed it.

Step 3: Graph-powered retrieval

Here’s where it gets interesting.

When you ask a question, Evidoc first identifies the entities in your question. Then it uses a proprietary graph algorithm to propagate relevance signals through the Knowledge Graph.

Think of it like ripples in a pond. Your question drops a stone at certain entity nodes. The relevance signal spreads outward through the graph’s connections — from entity to sentence, from sentence to related entity, from that entity to sentences in other documents.

The result: sentences that are relevant not just because they contain similar words, but because they’re connected through a chain of reasoning to your question.

Step 4: Multi-hop reasoning chains

This is where Evidoc catches what other tools miss.

Example: You ask “Do these invoices match the contract terms?”

The algorithm finds the contract clause: “The hourly rate for professional services shall be $120”
It follows the entity connection to the amendment: “Section 4.2 is hereby amended to reflect a rate of $150 per hour”
It then connects to the invoice: “Professional services: 40 hours × $120/hr = $4,800”

The invoice charges $120/hr, but the amendment changed the rate to $150/hr. This discrepancy spans three documents, and the Knowledge Graph algorithm found it by following the entity connections.

Each step in the chain is cited. You can click any citation to see the exact sentence highlighted on the original document.

Step 5: Relevance scoring and citation

After the graph retrieval surfaces candidate passages, a neural engine scores them for relevance to your specific question. The top results become the context for the AI’s answer.

Every sentence that contributed to the answer becomes a numbered citation. Click any number — the original PDF renders with that sentence highlighted at word-level precision.

Why this matters

The combination of Knowledge Graph reasoning and precision citations means:

Answers you can verify — every claim traces to a specific sentence in a specific document
Connections you’d miss — multi-hop reasoning finds relationships across documents that no amount of Ctrl+F would surface
Zero configuration — the Knowledge Graph builds automatically from your uploads

This is the engine behind every Evidoc answer. No hallucination can survive a click.

Try Evidoc Free →