Skip to main content

Semantic Knowledge Platform

Use this page when you need the public-safe explanation for how Helpifyr handles governed knowledge, context assembly, multilingual memory, and admitted model choice without creating a second truth system.

When to use this page​

  • You need to explain the semantic platform to an operator, integrator, or domain owner.
  • You need to understand how Bobbin, Fabric, Warp, Reed, and Dobby split responsibilities.
  • You need one owner-clear description of context graphs, memory lanes, and fail-closed model selection.

Prerequisites​

  • You understand the basic owner split between source domains, shared contracts, reviewed docs materialization, and runtime readback.
  • You can distinguish a source of truth from a projection or retrieval surface.

Owner model​

The semantic platform is intentionally split so no single runtime starts inventing business truth.

  • domain systems
    • own business facts, documents, identities, and operational state
  • Fabric
    • owns ontology, authority, model admission, extension admission, and fail-closed verdict semantics
  • Bobbin
    • projects reconstructable knowledge, assembles bounded context, and operates multilingual memory lanes
  • Warp
    • binds agents to admitted knowledge, context, and model profiles
  • Reed
    • is the read-only agent-visible gateway for context delivery
  • Dobby
    • remains proposal-only and may not mutate truth or admission

Knowledge model​

The platform separates four different things that are easy to confuse:

  • source facts
    • domain-owned truth such as customer, document, task, or approval state
  • semantic projection
    • the reconstructable graph and retrieval layer Bobbin derives from source events and approved artifacts
  • context graph
    • the small task-specific subset delivered to an agent or operator workflow
  • episodic memory
    • remembered interaction or session facts stored separately from domain-truth projection

Architecture / Flow​

Step-by-step procedure​

1. Start with owner truth, not with the graph​

The graph and memory layers are derived and must stay reconstructable. If a fact is disputed, the source owner still decides what is authoritative.

2. Assemble only bounded context​

The platform does not deliver one giant knowledge dump. It assembles only the nodes, edges, documents, and memory facts required for the current decision or workflow.

3. Treat episodic memory and semantic projection as different lanes​

Multilingual memory helps an agent or workflow recover previously captured interaction facts. It does not replace authority, contracts, or source-owned business truth.

4. Keep model choice admitted, not hardcoded​

The platform may use different admitted model profiles over time. Public docs must describe model choice as:

  • capability-bound
  • residency-aware
  • customer- or user-selectable when the profile is admitted
  • fail-closed when the requested model is blocked or unsupported

Do not document one permanent hardcoded LLM name as if it were the semantic platform itself.

5. Keep learning proposal-only​

Learning may recommend:

  • missing fact classes
  • context budget changes
  • source-priority adjustments
  • model-profile candidates
  • extension-onboarding improvements

It may not directly rewrite truth, expand permissions, or silently promote a new model or source.

Negative and degraded states​

Common fail-closed states include:

  • knowledge_missing
    • required fact or relation is unavailable
  • source_stale
    • the source or projection is older than the admitted freshness budget
  • authority_conflict
    • multiple sources disagree and no reviewed winner exists
  • model_not_admitted
    • a requested model or provider profile is not approved for this surface
  • degraded_retrieval
    • the semantic lane partially fell back, so a higher-risk action must stay blocked

Verification​

This page is being used correctly when:

  1. source truth, projection, context, and episodic memory are not treated as the same thing
  2. model choice is described as admitted and replaceable, not permanently hardcoded
  3. owner boundaries remain explicit after the explanation

Common failure modes​

Treating Bobbin as the business-truth owner​

Problem:

  • a reconstructable projection is misread as the canonical source.

Better path:

  • route disputed facts back to the domain owner and use Bobbin as a governed read model

Treating model access like execution authority​

Problem:

  • a reader assumes that seeing context or using a model also grants permission to act.

Better path:

  • keep authority, context delivery, and action execution as separate gates

Hardcoding one model name into the public truth​

Problem:

  • customer choice, provider retirement, or residency rules break the public explanation.

Better path:

  • describe admitted model profiles and fail-closed selection behavior instead

Source Truth​

Next paths​