Evaluating Self-Hosted LLM Platforms

Audience: technical leaders and engineers asking “Why not just use Open WebUI?”

This document explains the major capability categories used to evaluate self‑hosted LLM platforms (LibreChat, Open WebUI, Bionic GPT, Onyx). Each category explains what the user sees, what happens underneath, and—where relevant—what a tool definition typically looks like. A comparison matrix follows each section.

---

1. Interaction & Modalities

What this category covers

How users interact with models and which input/output modalities are supported beyond plain text.

What happens underneath

The platform must:

• Normalize user inputs into model‑specific request formats
• Manage binary assets (files, images) with lifecycle and access control
• Route requests to different model endpoints depending on modality

Platforms that only support chat are thin UIs. Platforms that support multiple modalities require orchestration, storage, and policy enforcement layers.

Typical tool definition (example)

web_search(query: string, recency_days?: number)
create_image(prompt: string, size?: string)
analyze_image(image_ref: File)
open_url(url: string)
list_attachments()
get_attachment(id:number, offset: number)

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
Text chat	✓	✓	✓	✓
File attachments (tool‑accessible)	Partial	Basic	✓	✓
Image generation	✓	✓	✓	✓
Image analysis / vision	Partial	Partial	✓	✓
Web search via tools	Partial	Limited	✓	✓
Deep research mode	✗	✗	✓	✓

---

2. Assistants, Agents & Projects

What this category covers

How systems define persistent AI behaviour and context.

What happens underneath

• Assistants are stored objects with system prompts, tools, and data bindings
• Projects/workspaces attach shared context across chats
• Agentic systems include planners, tool executors, and state machines

This is the dividing line between chat UIs and AI systems.

Here is project-level tool pseudocode that represents what a full platform (not just a chat UI) typically exposes to an LLM when operating inside a Project / Workspace.

Project Context & Navigation

get_project() -> Project
list_project_chats(project_id: string) -> Chat[]
get_chat(chat_id: string) -> Chat
create_chat(project_id: string, title?: string) -> Chat
rename_chat(chat_id: string, title: string)

Conversation Memory & State

store_memory(key: string, value: string)
read_memory(key: string) -> string
list_memory_keys() -> string[]
delete_memory(key: string)

File & Artifact Management

list_attachments() -> File[]
read_file(file_id: string) -> FileContent
write_file(name: string, content: bytes) -> File
delete_file(file_id: string)

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
Named assistants	✓	Limited	✓	✓
System instructions	✓	✓	✓	✓
Tool access per assistant	Limited	Limited	✓	✓
Projects / shared context	✗	Partial	✓	✓
Multi‑step agent execution	✗	✗	✓	✓

---

3. Knowledge & Data Ingestion (RAG)

What this category covers

How external knowledge is ingested, indexed, and retrieved during inference.

What happens underneath

• Documents are parsed and chunked
• Embeddings are generated and stored
• Retrieval pipelines select relevant chunks at runtime
• Permissions are enforced per chunk or source

This requires background pipelines, vector stores, and runtime binding—none of which exist in simple chat UIs.

Rag Tools passed to the model

search_knowledge(query: string, k?: number) -> RetrievedChunk[]
get_chunk(chunk_id: string) -> RetrievedChunk

request_full_document(
  source_id: string,
  justification: string
) -> DocumentHandle

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
RAG support	✓	✓	✓	✓
Multiple knowledge sources	Limited	Limited	✓	✓
Continuous ingestion	✗	✗	✓	✓
Permission‑aware retrieval	✗	✗	✓	✓
Supported file types (broad)	Partial	Partial	✓	✓

---

4. Integrations & Extensibility

What this category covers

How the system interacts with external APIs and tools.

What happens underneath

• OpenAPI specs are parsed into callable tools
• OAuth2 flows and secrets are managed centrally
• Tool execution is sandboxed and audited
• MCP enables cross‑system agent interoperability

This is where platforms stop being UIs and become integration hubs.

Typical OpenAPI tool binding

openapi: https://api.example.com/openapi.json
auth:
  type: oauth2
  token_ref: crm_access_token

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
OpenAPI tool ingestion	Partial	Partial	✓	✓
OAuth2‑secured APIs	✗	✗	✓	✓
MCP client	✗	✗	✓	✓
MCP server	✗	✗	✓	Limited
Virtual API keys	✗	✗	✓	✓

---

5. Security, Identity & Governance

What this category covers

Controls required in multi‑user, regulated, or enterprise environments.

What happens underneath

• Identity providers are federated via SSO
• Authorization is enforced at request and data levels
• Secrets and prompts are encrypted at runtime
• Audit events are emitted for compliance

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
SSO	Partial	Partial	✓	✓
RBAC	✗	✗	✓	✓
Teams / multi‑tenancy	Limited	Limited	✓	✓
Sharing controls	Basic	Basic	✓	✓
Runtime encryption	✗	✗	✓	✓

---

6. Observability & Cost Control

What this category covers

Visibility into usage, cost, and system behaviour.

What happens underneath

• Every request/response is metered
• Tokens are attributed to users, teams, and models
• Metrics are exported via OpenTelemetry

Without this, systems cannot be governed or scaled responsibly.

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
Token accounting	Partial	Partial	✓	✓
Request/response storage	Partial	Partial	✓	✓
OpenTelemetry support	✗	✗	✓	✓
Grafana‑ready metrics	✗	✗	✓	✓

---

7. Deployment & Platform Architecture

What this category covers

How the system runs in real infrastructure.

What happens underneath

• Services are decomposed and scalable
• Execution environments are isolated
• Kubernetes operators manage lifecycle and upgrades

This is the clearest separator between tools and platforms.

Comparison matrix

Capability	LibreChat	Open WebUI	Bionic GPT	Onyx
Self‑hosted	✓	✓	✓	✓
Kubernetes native	✗	✗	✓	✓
Kubernetes operator	✗	✗	✓	Partial
Sandboxed code execution	✗	✗	✓	✓

---

Final Takeaway

Open WebUI answers the question: “How do I chat with an LLM?”

Platforms like Bionic GPT and Onyx answer: “How do I safely, observably, and extensibility deploy AI across an organization?”

That distinction only becomes visible once you look below the UI layer—exactly what this evaluation framework is designed to do.