Skip to main content
ClipsX works without AI. An embedding model finds clips by meaning; a generation model writes Recall answers or runs compatible AI extensions. A chat model is not automatically an embedding model. ClipsX does not install Ollama or download models for you.

​
Start with the developer’s setup

The developer reports using bge-m3:latest for Meaning Search and gemma3:4b for generation. This is a starting example, not a performance guarantee for your device.
1

Install and start Ollama

Follow the official installation instructions for your OS. Start Ollama. Its default local endpoint is http://localhost:11434. If the service is already running, do not start a second copy.
2

Download the models

Open a terminal on the same device as ClipsX:
The list should contain the downloaded names. For search only, skip Gemma. Downloads need internet access; local inference uses the installed model.
3

Connect ClipsX

Open Intelligence → Models → New connection, choose Ollama, and enter http://localhost:11434. Save the connection and select Test and discover models. Refresh after downloading another model.
4

Enable the capabilities you want

Add a text-embedding preset for bge-m3:latest and a chat preset for gemma3:4b. In Usage & defaults, assign the embedding preset to Semantic search and the chat preset to Recall assistant or Extension text generation. Each feature has its own selection. Search validates the embedding configuration and indexes eligible clips in the background.
Continue with your first Meaning Search, your first Recall answer, or a Rewrite transformation. Exact text search remains available during semantic indexing.

​
Choose an embedding model

Embedding models produce vectors for retrieval; they do not write answers. ClipsX lists embedding-capable models reported by Ollama and validates their output when enabled. Alternatives are candidates to evaluate, not configurations independently benchmarked in ClipsX. Download one, refresh Models, and choose it under Meaning Search:
Changing embedding models changes the vector space. ClipsX builds a replacement index; during rebuilding, results can use the existing active index. Wait until the replacement is ready before comparing models.

​
Choose a generation model

Generation writes text. Recall can use image context and tools when the selected model and provider support them. Extension generation has its own capabilities and assignment; a model name alone does not establish support. For example:
Refresh the connection, add or update its chat preset, and assign it in Usage & defaults. Compare a Recall question with citations and a short Rewrite task. Assess usefulness, errors, waiting time, and resources—not only whether text appears. Changing a chat preset does not require rebuilding the semantic index. Listings reviewed October 8, 2026. Tags and requirements can change. Check linked official listings for sizes, minimum Ollama versions, and model licenses. Model licenses are separate from ClipsX’s Apache 2.0 license. When reporting a problem, record the exact tag and identifier shown by ollama list.

​
Understand resource costs

Review each model’s license and upstream terms: BGE-M3, Gemma 3, EmbeddingGemma, Gemma 4, Qwen 3.5, Qwen3 Embedding upstream. For Qwen models, follow the license or model-card links on those official pages. Sharing the Ollama API does not make model licenses identical. Download size and parameter count are not RAM or VRAM requirements. Runtime memory depends on quantization, context, and loaded models. CPU inference may be slower; larger contexts can require more memory. No fixed hardware tier or speed is promised. Use ollama ps during a task to inspect loaded models and CPU/GPU placement. Start with short text. If the device struggles, evaluate a smaller model, close heavy workloads, or use exact search without AI.

​
Keep inference local

Use locally downloaded tags. A loopback endpoint alone does not prove local inference: Ollama also offers cloud-backed models. Check the selected model and Ollama settings. Avoid cloud variants for device-local processing. Models receive the text needed for each task; extensions have their own permission boundaries.

​
Use a hosted provider

You can also add OpenAI, Anthropic, Gemini, or OpenRouter connections in Intelligence → Models. Save the provider’s credential, test the connection, and create a preset for a supported purpose. Assign it in Usage & defaults. Existing chats keep their own model selection. Hosted models receive the content needed for the selected task. Background clipboard indexing requires separate explicit permission. Provider capabilities vary; not every provider supports embeddings, images, or tools. Credentials stay in this device’s protected storage and are excluded from configuration exports. On macOS, Keychain may ask permission. Removing a connection clears its model assignments rather than silently selecting a replacement.

​
Recover or disable

  • Cannot connect: confirm Ollama is running and the endpoint matches. Check ollama list before reconnecting.
  • No suitable model: download one with the required capability, refresh, and choose its installed name.
  • Indexing fails: review Intelligence → Indexing, affected clips, and diagnostics before retrying.
  • Generation fails or answers poorly: check the installed model, try shorter input, and evaluate another model. The same API does not guarantee support or useful output.
  • Stop AI: clear the feature assignment in Usage & defaults or disable its connection. Removing stored semantic data is a separate index action; original clips remain.
See Troubleshooting and privacy boundaries.

​
Decision models

In ClipsX 0.1.8 and later, decision models evaluate yes/no questions about new copied text. General decisions serves organization rules, extension automation conditions, and permitted extension requests. Privacy checks separately serves credential classification and restrictive capture rules. Generation still writes the resulting text. Follow Decision model setup, then try a harmless tag rule or conditional Rewrite automation. Use the privacy action and failure guide before relying on a rule to restrict capture. Hosted providers receive evaluated text; local inference is recommended for privacy.