My Journey

CoPilot model config

3 min read

obsidian copilot plugin guidance

Understanding Obsidian Copilot Model Icons & Capabilities

The short answer is no, you do not need to disable web search for models displaying the web icon.

The behavior you experienced where a model claims it is strictly “text-based and cannot retrieve from a web page” is a common quirk of how AI models are prompted vs. how client-side plugins execute tools.

Decoding the UI Icons in Your Model Dropdown

In Obsidian Copilot, the small icons next to model names represent features and capabilities natively supported or routed through Copilot:

  • Lightbulb (💡): Indicates Reasoning / Thinking capability. Models with this icon (like nemotron-3-nano... or gpt-5-nano) output intermediate reasoning chains before returning an answer.

  • Eye (👁️): Indicates Vision / Multimodal capability. These models can process images, wireframes, or diagrams embedded in your notes.

  • Globe (🌐): Indicates Web Search / Browsing Capability. This shows the model endpoint or Copilot integration supports fetching live web content.

Why Models Say “I Cannot Access the Web”

When a model with a globe icon (like mistral-small-3.2-24b-instruct) tells you it cannot browse the internet, one of two things is happening:

1. System Prompt Hallucination (Most Common)

Most large language models are base-trained with system instructions like “I am a helpful AI assistant. I do not have access to live internet or real-time data.”

  • Unless Copilot passes a real-time web search snippet into the prompt before sending it to the model, or uses Agent Mode / Function Calling, the base model defaults to its core training and refuses to attempt a web lookup.

2. Copilot Mode Differences

Copilot processes external context differently depending on which mode you are using:

ModeWeb CapabilitiesHow It Works
Standard ChatPassive Context FetchingIf you paste a URL into the chat, Copilot scrapes the web page or YouTube transcript and feeds the raw text into the model’s context window.
Agent Mode (Plus)Autonomous Function CallingThe model calls a web search API automatically when it detects a prompt requiring external information.

Relevant Content (General Guidance)

Should You Disable Web Search on Custom Models?

  • Keep Web Search Enabled: Leaving the globe option active allows Copilot to pass web-scraped context (URLs, live search queries) to the API endpoint.

  • Prompting Around the Refusal: If a model refuses a web request, explicit prompting helps bypass its base safety rules. For example:

    “Use the provided web context / URL to answer the following question…”

  • Model Choice for Web Tools: Smaller or instruction-tuned models (e.g., llama-3.2-3b-instruct or local models) often fail function-calling schema validation even when configured for tools. For reliable web searching and tool execution, models with explicit function-calling tuning (like gpt-4o/gpt-5 variants or gemini-2.5-flash) perform significantly better.

Personal Details & Custom Vault Workflow

Markdown Formatting for Obsidian Canvas & PKM

Per your note formatting requirements for Obsidian logging:

Markdown

\#\# Copilot Model Capabilities Reference

\- \*\*Vision Icon (👁️)\*\*: Vision processing (Images, PDFs, Diagrams)
\- \*\*Globe Icon (🌐)\*\*: Web Search / Context Retrieval enabled
\- \*\*Lightbulb Icon (💡)\*\*: Chain-of-thought / Reasoning mode

\#\# Recommended Model Selection by Task

1. \*\*General Note Chat & Summaries\*\*: google/gemini-2.5-flash (Fast, reliable function calling)
2. \*\*Complex Reasoning / Math / Logic\*\*: nvidia/nemotron-3-nano-omni-30b-a3b-reasoning
3. \*\*Local / Privacy Workflows\*\*: mistralai/mistral-small-3.2-24b-instruct

Graph View