Day-0 Guide
General enterprise visibility, with access to the enterprise's general document spaces. No HR or payroll compartment access.Viewing as Enterprise Reader
Chapter 11 of 13EKA V1.1 DIRECTION

What happens when an LLM can use both enterprise facts and enterprise knowledge?

By this point, Metrolane has two important foundations.

Structured enterprise knowledge:

canonical entitiesbusiness objectsrelationshipsscenariosprovenance

and document knowledge:

policiesmanualscontractsreportsretrieval-ready passages

EKA — the Enterprise Knowledge Assistant — is one solution built on top of those foundations.

Understand

EKA is not Metrolane itself

Metrolane is the enterprise environment.

EKA is a consumer of that environment.

Metrolane enterprise

+

Metrolane document retrieval

consumed by

EKA

produces

LLM-grounded answer

Two kinds of context

A useful enterprise question may need structured facts.

What is the status of this warranty claim?Which policy reference does it use?What decision was recorded?

It may also need document knowledge.

What does the applicable warranty policy say?

EKA can eventually bring those together.

See it in Metrolane

Warranty claim + policy

Structured facts
Warranty ClaimWCL-2026-000002
  • claim status = REJECTED
  • claim reason = bracket bent after a kerb strike
  • policy reference = WP-EXT-36
  • decision = REJECTED
Document knowledge
Extended 36-Month Warranty PolicyDOC-000050

The document estate includes the retrieval-ready policy.

Candidate teaching question:

Why was warranty claim WCL-2026-000002 rejected, and what policy applies to it?

RAG in enterprise terms

A simple RAG diagram often looks like:

question → retrieve chunks → LLM

For enterprise use, the real shape is richer.

user question

reader identity / access context

structured enterprise context+eligible / applicable document retrieval

LLM reasoning

grounded response

sources / provenance

The retrieval layer should not ignore enterprise access or context simply because a text chunk has high semantic similarity.

Important distinction

Semantic Search vs EKA

Day-0 Semantic Searchfind relevant document passages
EKAcombine enterprise context + retrieved knowledge + LLM reasoning

The first is a retrieval experience.

The second is an LLM-driven solution.

Current architecture boundary

The current Metrolane estate already has reusable structured-read and document-retrieval foundations.

The final EKA design still needs to freeze details such as:

exact tool/query catalogueorchestration flowanswer contractcitation presentationfailure behaviorstructured + unstructured context composition

The Guide should not invent those details before the EKA architecture is finalized.

Important distinction

Access must survive the AI layer

If a normal Enterprise Reader cannot retrieve an HR document, EKA must not use that same document simply because the LLM would find it useful.

EKA accessmust not be weaker thanunderlying enterprise/retrieval access
See it in Metrolane

Access-aware grounding

Offer Letter — Devika IyerDOC-000070

Under different reader contexts, the visible retrieval universe changes.

This does not yet demonstrate a final EKA answer UI. It demonstrates the grounding boundary EKA must respect.

Go deeper

Structured + unstructured grounding

Structured + unstructured grounding

The eventual EKA architecture should be able to distinguish:

  • facts asserted in canonical data
  • facts asserted in source data
  • retrieved document passages
  • scenario/read-model context
  • provenance

The LLM should reason over those inputs without silently collapsing their authority.

For example:

Recorded claim decision
policy text
model inference

The final answer experience should make those distinctions understandable.

Chapter takeaway

  • EKA is one LLM-driven solution built over Metrolane.
  • It can combine structured enterprise context with governed document retrieval.
  • Semantic Search retrieves.
  • EKA reasons over enterprise and retrieval context.
  • Access and provenance must survive the LLM layer.