Your business, readable by an agent.

Components

What runs in this layer.

Context index · retail demo tenant

Demo data

SourceFreshness

OrdersSAP · MCP · entitylive
CustomersSalesforce · MCP · entitylive
Return policy, v7SharePoint · 148 passages2 h ago
Carrier SLAs by regionS3 · 62 passagesdaily
Past casesZendesk · 3,410 casesreindexing

Every passage keeps its source. The index lives in your AWS account.

07 components in this layer

In production× 6
Reference architecture× 1

01

Entity model

Typed entities with an owner system and a source.

02

Business rules

Limits, escalations and thresholds, versioned like code.

03

Amazon Bedrock Knowledge Bases

Managed ingestion and retrieval in your AWS account.

04

pgvector and OpenSearch

Vector and keyword indexes, for part numbers and paragraphs.

05

Hybrid RAG with reranking

A reranker picks what actually reaches the model.

06

Short and long-term memory

This conversation and the last one, with your retention rules.

07Reference architecture

Neo4j knowledge graph

Multi-hop questions where relations beat documents.

In production: A real client, or DaCodes itself labeled as internal use, with a KPI measured in the Proof Ledger.

The model of your business

What an agent has to know.

01

Entities

Each with an owner system and a freshness rule.

02

Rules

What the agent may decide alone, and what needs a person. Versioned like code.

03

Documents

Manuals, contracts and past cases, indexed with the source on every passage.

04

Memory

This conversation and the last one, with retention you control.

Call it an ontology, a semantic layer or a knowledge graph. We call it the context layer: the part that makes the agent yours.

context/order · #48213

Demo data

Order #48213

live

01Entities

status: Failed delivery, retry today
customer: C-20931
total: $2,480.00 MXN
store: Store 112, Monterrey

Linked entities

Customer
Shipment
Carrier
Store
Return

02Rules

Refund up to $1,500 MXN alone
Escalate on the 2nd failed deliveryapplies

03Documents

Return policy, v7SharePoint · 148 passages

04Memory

09:41WhatsApp: asked where the order is; told the carrier retries today.

Read over MCP. Evals rerun when a rule changes.

Anonymized example

One retail agent, drawn as its context.

Six entities, four rules, two document sets. It resolves the routine questions alone and escalates the rest with the evidence.

Context graph · order-status agent

Anonymized

Rules · 4

Refund up to $1,500 MXN without approval
Escalate on the second failed delivery
Never promise a date the carrier has not confirmed
Returns only within 30 days of delivery

Documents · 2

Return policy, current version
Carrier SLAs by region

Memory · 2

Last three contacts with this customer
Open case and promised follow-ups

Composite drawn from retail deployments in México; names and thresholds changed. Sep 2026.

Questions

Architects ask.

Close, without the ceremony. The context layer holds typed entities, relations and rules like an ontology would, but it starts from the six or seven entities your first agent needs and grows with each one. You never model the whole company up front.

In your account. Indexes run on pgvector, OpenSearch or Bedrock Knowledge Bases inside your AWS tenant; the source documents stay where they are. Nothing is copied to a DaCodes environment.

Each source has a sync schedule and a freshness rule. When a policy changes, the index updates and the eval set that depends on it runs again, so you see whether the agent's answers moved.

Talk to an architect

45 minutes with the people who run the platform. No slides.

We reply in under one business day.