Choose your models
Connect Ollama locally or configure a supported cloud provider. Your choice changes hardware needs, cost and where prompts are processed.
Open-source AI search · Docker + Ollama
Vane: an open source Perplexity alternative you can self-host
Build a search-and-answer workflow with your own models and storage. Compare Vane with Perplexity, then follow a local Docker setup with Ollama.
Independent guide, not a hosted search service. Installation examples are documented instructions, not a claim that we ran or benchmarked Vane.
Sources checked:
What is Vane?
Vane combines a search backend and language models to produce answers with source links. The application is MIT-licensed; the models and services you connect have their own terms.
Search with source links
Vane uses SearXNG for web retrieval. Open cited pages to check whether they support the answer; citations do not guarantee accuracy.
Own the running service
Keep application data in a persistent volume and control updates and access. You also take responsibility for storage, backups and availability.
Vane vs Perplexity
A workflow comparison based on official documentation, not an accuracy or speed benchmark. Features and limits can change; check the linked sources before choosing.
| Dimension | Vane · self-hosted | Perplexity · hosted service |
|---|---|---|
| Source citations | Search-backed answers with source links; verify the linked pages. | Web-first answers with citations; verify the linked pages. |
| Model choice | Configure local Ollama or a supported provider with your own credentials. | Models and selection options depend on the current product and plan. |
| Self-hosting | Run the MIT-licensed application on infrastructure you control. | Use the hosted product; this comparison does not describe a self-hostable Perplexity edition. |
| File questions | Document uploads; processing depends on the configured models and file support. | File uploads and analysis with plan-dependent limits. |
| Data flow | You control the app and storage. Search requests still leave the machine; cloud models add another processor. | Requests are processed by a hosted service under its product policies and account settings. |
| Maintenance | You manage updates, model downloads, backups and network access. | The vendor operates the service; no local inference stack to maintain. |
| Cost | No application license fee under MIT; budget for hardware, electricity, hosting and any paid providers. | Free and paid plans with different limits. Check current terms rather than assuming unlimited access. |
What self-hosting does—and where data still travels
This is a simplified logical flow. Vane orchestrates retrieval and model calls; SearXNG does not directly become the language model.
- 01
Your question
- 02
Vane
- 03
SearXNG retrieval
- 04
Local or cloud model
- 05
Answer + citations
Local inference
Select locally downloaded models for chat and, when needed, embeddings. Inspect all configured providers; selecting Ollama alone does not prove every step is local. Ollama also offers cloud features, which can be disabled in its settings.
Web retrieval
A web search needs network access. Search engines and visited sites may receive queries, derived searches and connection metadata even when the answer is generated locally.
Cloud inference
If you choose a cloud provider, prompts and relevant context go to that provider. Review its retention and billing terms before sending documents or sensitive material.
Keep this tutorial local. Before sharing an instance, put HTTPS and authentication/access control in front of it. Protect model endpoints too; opening a port is not an access policy.
Self-host Vane with Docker + Ollama
Install Docker and Ollama on your own machine. The commands below are examples to run there; this website does not install software or download models for you.
- 01
Prepare Docker and Ollama
Start Docker and the Ollama service. Match model size to available memory and disk space. CPU inference is possible but can be slow; no fixed GPU or RAM requirement is promised here.
- 02
Download example models
Pull llama3.2:3b for a small text-chat starting point. nomic-embed-text is an embedding model for document retrieval, not a chat model. These are examples, not an equivalence claim against commercial systems.
- 03
Start the bundled image
The official full image includes SearXNG. This guide maps the web app to 127.0.0.1:3001 and stores data in vane-data. Open http://localhost:3001 after startup.
- 04
Configure and search
In Vane’s setup, add Ollama using the address for your OS below, select the downloaded chat model and configure the embedding model for file retrieval when needed. Ask a factual question, then open its citations and check the source text.
Download and list models
ollama pull llama3.2:3b ollama pull nomic-embed-text ollama list
Start Vane · local web access only
docker run -d --name vane --restart unless-stopped -p 127.0.0.1:3001:3000 -v vane-data:/home/vane/data itzcrazykns1337/vane:latest
Commands shown as single lines work in PowerShell and bash/zsh. The example uses a mutable latest tag for the introductory path; review release notes and pin an image version or digest for repeatable deployments.
Connect the container to Ollama
Inside a Docker container, localhost refers to that container, not the host machine. Use the correct host address and confirm the Ollama listener is reachable.
Windows / macOS · Docker Desktop
Use http://host.docker.internal:11434 in Vane. If unreachable, follow the Ollama FAQ to configure its listener and restart the app. Restrict network access to the Docker host/container network.
Linux · host networking details
Use http://<HOST_PRIVATE_IP>:11434 with your actual reachable host address. A host-gateway DNS mapping alone cannot reach a service bound only to 127.0.0.1. Configure the listener and firewall before retrying.
Run sudo systemctl edit ollama.service, enter the override below, save, then run the reload commands separately. Binding 0.0.0.0 listens on all interfaces: allow only the Docker/private network in your firewall and never forward port 11434 to the public internet. Windows and macOS use different environment-setting steps; follow the official FAQ.
Linux systemd override · edit, then reload
[Service] Environment="OLLAMA_HOST=0.0.0.0:11434"
Reload the Linux Ollama service
sudo systemctl daemon-reload sudo systemctl restart ollama
Troubleshooting checklist
Start with the model service and container logs, then check the search backend.
- No downloaded model appears
- Run ollama list on the host, verify the model tag and confirm Vane reaches the same Ollama service. A successful download into another installation is not enough.
- Cannot connect to Ollama
- Check the OS-specific host address, port, listener and firewall. Container localhost is not the host. Do not solve this by exposing the model API publicly.
- Port or container name is already in use
- Inspect docker ps -a. Use another unused host port while keeping the container port 3000, or stop the old instance deliberately. Do not remove unrelated containers.
- Search returns no sources
- Inspect the Vane logs, try a simpler public query and verify SearXNG can reach its configured engines. For a separate SearXNG instance, check its URL and JSON support.
- Answers are slow or run out of memory
- Try a smaller model or shorter context, check available memory and disk, and review Ollama logs. A lightweight model is a starting point, not a guarantee of high-quality research.
Keep data, inspect logs and update
The named volume vane-data is mounted at /home/vane/data. It outlives a replaced container; deleting the volume deletes its contents. Back it up before upgrading.
Choose a fresh backup directory instead of overwriting an earlier backup. The example pauses Vane while copying. Backups may contain configuration, credentials, uploads and conversations; store them privately. Keep the backup and record the current image digest before updating.
Inspect container status and logs
docker ps -a --filter name=vane docker logs --tail 100 vane
Stop, copy a backup, restart
docker stop vane docker cp vane:/home/vane/data ./vane-backup docker start vane
Update while reusing the data volume
docker pull itzcrazykns1337/vane:latest docker stop vane docker rm vane docker run -d --name vane --restart unless-stopped -p 127.0.0.1:3001:3000 -v vane-data:/home/vane/data itzcrazykns1337/vane:latest
The update removes only the stopped container, then starts the new image with the same volume and local port binding. Do not delete vane-data. After updating, check settings, run a search and inspect citations; rollback may also require restoring data compatible with the old version.
Other configuration paths
Use a cloud model
Choose a supported cloud provider in Vane’s setup and supply your own API key. Cloud usage has separate billing and data handling. Keep keys in the instance configuration; never submit them to this website.
Use an existing SearXNG instance
The slim image uses your existing SearXNG URL via SEARXNG_API_URL. The official guide requires JSON format and the Wolfram Alpha engine to be enabled. Use a URL reachable from the Vane container, not its localhost. Follow the linked instructions for the full command.
Which one fits your workflow?
Choose Vane for control
Advantages: configurable local models, inspectable application code and control over app storage. Tradeoffs: installation, model quality, slower local inference and ongoing maintenance are your responsibility.
Choose Perplexity for convenience
Advantages: a managed search experience without running a model stack. Tradeoffs: service-defined models, plan limits and hosted data processing. Compare current features against your actual research tasks.
This is a fit assessment, not a benchmark. A small local text model does not inherit every capability shown in Vane’s product description; image understanding and research quality depend on the model you configure.
Frequently asked questions
Official references
Features and installation details are based on the sources below, checked on the stated date. Consult the latest release instructions before running the commands.