ENGINEERING

04 / CONTEXT ENGINEERING

Relevant context.
Better-founded answers.

An AI system needs more than a prompt. We design how it finds, filters and uses information, with the permissions, freshness and evidence your application requires.

Discuss your engineering priorities ARCHITECTURE THROUGH TO OPERATION

SYSTEM VIEW

Build an information path you can evaluate.

Context should be selected for the task and the caller. Its origin, access rules and limits remain visible to the application.

Permission-aware context assembly

  1. 01

    Establish the task

    Resolve user identity, access scope and the information the task requires.

  2. 02

    Retrieve & filter

    Search authorised sources and remove irrelevant or out-of-scope material.

  3. 03

    Assemble

    Prepare instructions, evidence, tool outputs and necessary state within a defined budget.

  4. 04

    Check the result

    Evaluate groundedness, citations and task completion, with an abstention path.

Context sources governed by the same access boundary

  • RETRIEVAL

    Document evidence

    Relevant, current source material with access filtering and traceable provenance.

  • INTEGRATION

    Live tool context

    Bounded results from authorised systems, validated before use.

  • CONTINUITY

    Conversation state

    Task-relevant history with retention, compaction and isolation rules.

Track retrieval quality, context size, source freshness and answer quality using minimised traces.

Illustrative context architecture. Retrieved content is treated as data; application permissions and tool approvals remain enforced outside the model.

WHAT WE ENGINEER

Useful information.
Clear operating boundaries.

Retrieval design

Shape ingestion, search, filtering and reranking around actual questions. Preserve source metadata and test whether the needed evidence is retrieved.

Permissions and provenance

Apply source-level access policies before information reaches the model. Make evidence and freshness available where users need to judge an answer.

Tool and state integration

Connect bounded tool results and scoped conversation state. Define what can persist, what expires and what requires a new read from the source system.

Evaluation

Create representative cases for retrieval and answer quality, including missing evidence, conflicting sources and malicious instructions in retrieved content.

CONTEXT WITH ACCOUNTABILITY

Keep the model informed.
Keep authority in your systems.

We connect retrieval and application controls so a model receives information appropriate to the task. Context limits, source changes and tool behaviour are tested as part of the integration.

Adding documents does not guarantee a correct answer. We define what evidence is sufficient, how an answer can be checked and when the system should ask for clarification or decline to answer.

Anthropic: effective context engineering

ENGINEERING QUESTIONS

Decisions worth getting right.

Is context engineering the same as RAG?

Retrieval-augmented generation is one part of the context design. The wider system also includes instructions, authorised tool outputs, conversation state, information limits and the evaluation of what the model does with them.

Can users see information they cannot access in the source?

They should not. We enforce retrieval and tool permissions in the application and source integrations, rather than relying on a model instruction to hide information after it has been retrieved.

Will this eliminate incorrect answers?

No. Better information can improve a system, but output still needs evaluation and a defined response when evidence is missing or contradictory. We establish quality criteria around the actual workload.

LET’S ENGINEER WHAT COMES NEXT

Give your AI
a better information foundation.

Bring us a workload, a technical constraint or an architecture that needs a second look. We will help define a practical next step.

Discuss your engineering priorities

Architecture advice, focused implementation and support for your engineering team.