Synbuild

PKA architecture

The LLM generates language; the PKA controls the working environment

A good Personal Knowledge Assistant is distinguished by how it selects knowledge, applies a method, respects access and makes outputs reviewable—not by claiming to own a unique model.

Reviewed
On this page
  1. What still sits outside the LLM
  2. Configuration and governance remain necessary around the model
  3. Nine context types can influence one task
  4. More context does not automatically produce a better answer
  5. Synbuild does not rebuild the complete AI stack at every level
  6. Standard platform capability will continue to change the lightest viable PKA

The complete layer

What still sits outside the LLM

Each layer constrains or enriches the next. The exact implementation can use an existing AI product, a repository, retrieval services or a custom application.

  1. Layer 1

    User

    The person, role, preferences, permissions and responsibility for review.

  2. Layer 2

    Interface

    The place where questions, files, results, citations and approvals are presented.

  3. Layer 3

    Instructions and agent roles

    Tone, method, responsibilities, prohibited behaviour and output formats.

  4. Layer 4

    Workflow and orchestration

    Task steps, tool order, checkpoints, escalation and hand-offs.

  5. Layer 5

    Retrieval and knowledge selection

    Which authoritative and permitted context is relevant to this question and task.

  6. Layer 6

    Rights, logging and approval

    Access checks, audit records, evaluations, human approval and error handling.

  7. Layer 7

    LLM

    The selected model interprets context and generates or structures an output.

  8. Layer 8

    Tools, databases and business systems

    Search, storage, CRM, ERP, document platforms and bounded external actions.

PKA = LLM + context selection + memory + workflows + control

The distinguishing work

Configuration and governance remain necessary around the model

Source authority
Determine which documents govern an answer and how they are added, replaced or removed.
Knowledge status
Separate drafts, reviewed material, approved knowledge and deprecated versions.
Versioned method
Keep instructions, prompts, roles and workflows reviewable and exportable.
Access and approval
Apply rights by user or department and require approval for sensitive outputs or actions.
Evaluation and logging
Test representative questions, inspect citations and record errors and usage.
Integration and portability
Connect systems where justified and preserve the ability to change model providers.

Context is broader than prompt text

Nine context types can influence one task

Instruction context
Tone, attention points, required citations and analysis priorities.
Knowledge context
Relevant passages retrieved from selected business or personal documents.
User context
Role, preferences, expertise and permissions.
Task context
The result that must be achieved now.
Process context
Steps, checks, approvals and escalation rules.
Historical context
Relevant earlier conversations, decisions and approved outputs.
System context
Available tools, applications, schemas and data access.
Operational context
Current customer, stock, financial or project information.
Governance context
What is allowed, prohibited or possible only after approval.

Why not send everything?

More context does not automatically produce a better answer

Relevance
Thousands of pages contain information that is unrelated to the current question.
Contradiction
Old and current policies can provide incompatible instructions.
Cost and speed
Sending complete collections repeatedly is inefficient and can reduce focus.
Security
A user may be entitled to only part of the available business knowledge.

Consequences for the four levels

Synbuild does not rebuild the complete AI stack at every level

Personal
Mainly analysis, instructions, workflows, templates and configuration in a suitable existing platform.
Professional
Configuration plus a managed, exportable knowledge and file structure.
Knowledge
Controlled retrieval only when built-in file features lack scale, source control or measurable quality.
Business
A separate application layer for users, rights, integrations, logging, approval, monitoring and management.

Model providers are adding more of the stack

Standard platform capability will continue to change the lightest viable PKA

OpenAI currently documents Projects that combine chats, uploaded files, project instructions and project memory, as well as saved memory and chat-history references. Anthropic documents Claude Projects with project knowledge and retrieval, plus API tools for memory and active context management. These features mean that simple PKA functionality can increasingly be configured in standard products. They do not automatically settle an organisation’s source authority, evaluation, rights, portability, integrations or operating model. Future provider development is an expectation, not a guarantee; features and plan limits can change.

Sources

Official product documentation

Last reviewed: 26 July 2026

  1. Projects in ChatGPT — OpenAI
  2. How reference saved memories work — OpenAI
  3. What are projects? — Anthropic
  4. Context editing — Anthropic
  5. Memory tool — Anthropic

The descriptions above are limited to these official sources as reviewed on the stated date. Availability, behaviour and subscription requirements can change.