Self-Host Perplexica: Open Source Perplexity AI in 2026

TL;DR

Perplexica is the closest open-source equivalent to Perplexity AI. It combines SearXNG (private web search) with any LLM backend — local Ollama models or cloud providers — to give you cited, AI-generated answers to your questions without sending data to Perplexity's servers. Self-hosting takes 10 minutes with Docker Compose and costs nothing to run if you already have a server.

Key Takeaways

Perplexica searches the web via SearXNG, reranks results with embeddings, and generates cited answers using your LLM of choice
Fully private: no queries leave your infrastructure when using Ollama + SearXNG
LLM flexibility: supports Ollama (local), OpenAI, Anthropic Claude, Google Gemini, Groq
Search modes: Web, Academic, YouTube, Reddit/Discussions, Wolfram Alpha
License: MIT — fully open-source, no cloud dependency required
Requirements: 2GB+ RAM for the stack; 8GB+ if running local Ollama models
Perplexity Pro costs $20/month — Perplexica costs $5–10/month in VPS fees or runs free on existing hardware

What Is Perplexica?

Perplexity AI became popular by combining real-time web search with LLM-generated answers and source citations. Rather than a keyword search engine returning links, it answers questions directly with inline citations you can verify.

Perplexica replicates this workflow entirely with open-source components:

User query
    ↓
SearXNG (private meta-search engine)
    ↓ top N results
Embedding reranker (filters most relevant results)
    ↓ relevant context
LLM (Ollama/OpenAI/Claude/Gemini)
    ↓
Cited answer with source links

The key difference from just using an LLM: Perplexica fetches current web content before generating the answer. Ask about a news event from yesterday and it finds the relevant articles first, then synthesizes a response — the LLM isn't relying on training data.

What You Get vs. Perplexity Pro

Feature	Perplexity Pro ($20/mo)	Perplexica (self-hosted)
AI-powered web search	✅	✅
Source citations	✅	✅
Local/private queries	❌	✅
Model choice	Limited	Any (Ollama, OpenAI, Claude…)
Academic search	✅	✅
Reddit/discussion search	✅	✅
YouTube search	✅	✅
Image generation	✅	❌
API access	✅ (paid)	✅ (self-hosted)
Mobile app	✅	❌
Monthly cost	$20	$0–10

The main gaps vs. Perplexity Pro: no mobile app, no image generation in the base install. Everything else the core search-and-answer workflow is replicated.

System Requirements

Minimum (cloud LLM like OpenAI/Groq):
  RAM: 2GB
  CPU: 1 vCPU
  Storage: 5GB

Recommended (local Ollama models):
  RAM: 16GB+ (for 7B models like Llama 3.1)
  GPU: Optional but recommended (NVIDIA for CUDA)
  Storage: 20GB+ (model weights)

Tested on:
  - Ubuntu 22.04 / Debian 12
  - VPS: Hetzner CPX21 (~$7/mo) for cloud LLM mode
  - Local: M-series Mac, modern Linux desktop for Ollama mode

Docker Compose Install

Perplexica ships as a multi-container Docker Compose stack. SearXNG is included automatically — you don't install it separately.

Step 1: Clone and Configure

git clone https://github.com/ItzCrazyKns/Perplexica.git
cd Perplexica
cp sample.config.toml config.toml

Edit config.toml:

[GENERAL]
PORT = 3000                    # Web UI port
SIMILARITY_MEASURE = "cosine"  # or "dot_product"

[API_KEYS]
OPENAI = "sk-..."              # Optional — for OpenAI models
GROQ = "gsk_..."               # Optional — for Groq (fast free tier)
ANTHROPIC = ""                 # Optional — for Claude

[API_ENDPOINTS]
OLLAMA = "http://ollama:11434" # If running Ollama in Docker
SEARXNG = "http://searxng:8080"

[CHAT_MODEL]
PROVIDER = "ollama"            # or "openai", "anthropic", "groq"
MODEL = "llama3.1:8b"          # Model name for your provider
CUSTOM_OPENAI_BASE_URL = ""    # For OpenAI-compatible endpoints
CUSTOM_OPENAI_KEY = ""

[EMBEDDING_MODEL]
PROVIDER = "local"             # Uses local HuggingFace model
MODEL = "togethercomputer/m2-bert-80M-8k-retrieval"

Step 2: Docker Compose

# docker-compose.yml — full Perplexica stack
version: "3.8"

services:
  perplexica-backend:
    image: itzcrazykns/perplexica-backend:main
    depends_on:
      - searxng
    ports:
      - "3001:3001"
    volumes:
      - ./config.toml:/home/perplexica/config.toml
      - perplexica-uploads:/home/perplexica/uploads
    networks:
      - perplexica-network
    restart: unless-stopped

  perplexica-frontend:
    image: itzcrazykns/perplexica-frontend:main
    depends_on:
      - perplexica-backend
    ports:
      - "3000:3000"
    environment:
      NEXT_PUBLIC_API_URL: http://localhost:3001
      NEXT_PUBLIC_WS_URL: ws://localhost:3001
    networks:
      - perplexica-network
    restart: unless-stopped

  searxng:
    image: searxng/searxng:latest
    volumes:
      - ./searxng:/etc/searxng
    networks:
      - perplexica-network
    restart: unless-stopped

networks:
  perplexica-network:
    driver: bridge

volumes:
  perplexica-uploads:

# Start the full stack
docker compose up -d

# Check all containers are running
docker compose ps

# View logs if something fails
docker compose logs -f perplexica-backend

Open http://localhost:3000 — the search UI is ready.

Step 3: Configure SearXNG

SearXNG needs its settings file to allow JSON format (required by Perplexica):

# Create SearXNG config directory
mkdir -p searxng

# Create settings.yml
cat > searxng/settings.yml << 'EOF'
use_default_settings: true

server:
  secret_key: "$(openssl rand -hex 32)"
  limiter: false
  image_proxy: true

search:
  safe_search: 0
  autocomplete: ""
  default_lang: "en"

outgoing:
  request_timeout: 10.0

engines:
  - name: google
    engine: google
    shortcut: g
  - name: bing
    engine: bing
    shortcut: b
  - name: duckduckgo
    engine: duckduckgo
    shortcut: ddg

ui:
  static_use_hash: true

enabled_plugins:
  - 'Hash plugin'
  - 'Search on category select'
  - 'Tracker URL remover'

formats:
  - html
  - json  # Required by Perplexica
EOF

Adding Ollama for Local Models

For fully private search (no data leaves your server), add Ollama to the stack:

# Add to docker-compose.yml services:
  ollama:
    image: ollama/ollama:latest
    volumes:
      - ollama_data:/root/.ollama
    ports:
      - "11434:11434"
    networks:
      - perplexica-network
    restart: unless-stopped
    # For GPU support, add:
    # deploy:
    #   resources:
    #     reservations:
    #       devices:
    #         - driver: nvidia
    #           count: all
    #           capabilities: [gpu]

volumes:
  ollama_data:  # Add this alongside perplexica-uploads

# Pull a model after starting Ollama
docker exec -it perplexica-ollama-1 ollama pull llama3.1:8b

# For a smaller, faster model on 4GB RAM:
docker exec -it perplexica-ollama-1 ollama pull mistral:7b

# Update config.toml to use Ollama:
# PROVIDER = "ollama"
# MODEL = "llama3.1:8b"
# OLLAMA = "http://ollama:11434"

With this configuration, your search queries go to your own SearXNG instance and answers are generated by your local Llama 3.1 — nothing reaches external services.

Search Modes

Perplexica ships with specialized search modes beyond general web search:

Web (default)
  → SearXNG web results → LLM answer with citations

Academic
  → Searches arXiv, Semantic Scholar, PubMed
  → Best for research questions

YouTube
  → Searches YouTube, returns video links with summaries
  → Useful for "how to" questions

Reddit / Discussions
  → Searches Reddit for community discussion
  → Good for "what do people think about X" questions

Wolfram Alpha (optional, requires API key)
  → Computational/factual queries
  → Math, unit conversions, factual lookups

Configure the default mode in config.toml:

[FOCUS_MODE]
DEFAULT = "webSearch"  # or "academicSearch", "youtubeSearch", "redditSearch"

Reverse Proxy with Caddy

For public access with HTTPS:

# Caddyfile
search.yourdomain.com {
    reverse_proxy localhost:3000
}

# With Caddy installed:
caddy run --config /etc/caddy/Caddyfile

# Or add to docker-compose.yml as a Caddy service

Cost Comparison

Setup	Monthly Cost	Privacy	Speed
Perplexity Pro	$20	❌ Cloud	⚡ Fast
Perplexica + Groq (free tier)	$5 (VPS only)	⚠️ Groq API	⚡ Fast
Perplexica + OpenAI	$5–15 (VPS + API)	⚠️ OpenAI API	⚡ Fast
Perplexica + Ollama (cloud)	$15–30 (GPU VPS)	✅ Full	⚡ Moderate
Perplexica + Ollama (local)	$0 (existing hw)	✅ Full	⚡ Depends on GPU

The sweet spot for most developers: run Perplexica on a $5/month VPS with Groq's free API tier (28K tokens/minute free) for fast, nearly free AI-powered search with only the LLM call leaving your server.

Why Self-Host Perplexica?

The case for self-hosting Perplexica comes down to three practical factors: data ownership, cost at scale, and operational control.

Data ownership is the fundamental argument. When you use a SaaS version of any tool, your data lives on someone else's infrastructure subject to their terms of service, their security practices, and their business continuity. If the vendor raises prices, gets acquired, changes API limits, or shuts down, you're left scrambling. Self-hosting Perplexica means your data and configuration stay on infrastructure you control — whether that's a VPS, a bare metal server, or a home lab.

Cost at scale matters once you move beyond individual use. Most SaaS equivalents charge per user or per data volume. A self-hosted instance on a $10-20/month VPS typically costs less than per-user SaaS pricing for teams of five or more — and the cost doesn't scale linearly with usage. One well-configured server handles dozens of users for a flat monthly fee.

Operational control is the third factor. The Docker Compose configuration above exposes every setting that commercial equivalents often hide behind enterprise plans: custom networking, environment variables, storage backends, and authentication integrations. You decide when to update, how to configure backups, and what access controls to apply.

The honest tradeoff: you're responsible for updates, backups, and availability. For teams running any production workloads, this is familiar territory. For individuals, the learning curve is real but the tooling (Docker, Caddy, automated backups) is well-documented and widely supported.

Server Requirements and Sizing

Before deploying Perplexica, assess your server capacity against expected workload.

Minimum viable setup: A 1 vCPU, 1GB RAM VPS with 20GB SSD is sufficient for personal use or small teams. Most consumer VPS providers — Hetzner, DigitalOcean, Linode, Vultr — offer machines in this range for $5-10/month. Hetzner offers excellent price-to-performance for European and US regions.

Recommended production setup: 2 vCPUs with 4GB RAM and 40GB SSD handles most medium deployments without resource contention. This gives Perplexica headroom for background tasks, caching, and concurrent users while leaving capacity for other services on the same host.

Storage planning: The Docker volumes in this docker-compose.yml store all persistent Perplexica data. Estimate your storage growth rate early — for data-intensive tools, budget for 3-5x your initial estimate. Hetzner Cloud and Vultr both support online volume resizing without stopping your instance.

Operating system: Any modern 64-bit Linux distribution works. Ubuntu 22.04 LTS and Debian 12 are the most commonly tested configurations. Ensure Docker Engine 24.0+ and Docker Compose v2 are installed — verify with docker --version and docker compose version. Avoid Docker Desktop on production Linux servers; it adds virtualization overhead and behaves differently from Docker Engine in ways that cause subtle networking issues.

Network: Only ports 80 and 443 need to be publicly accessible when running behind a reverse proxy. Internal service ports should be bound to localhost only. A minimal UFW firewall that blocks all inbound traffic except SSH, HTTP, and HTTPS is the single most effective security measure for a self-hosted server.

Backup and Disaster Recovery

Running Perplexica without a tested backup strategy is an unacceptable availability risk. Docker volumes are not automatically backed up — if you delete a volume or the host fails, data is gone with no recovery path.

What to back up: The named Docker volumes containing Perplexica's data (database files, user uploads, application state), your docker-compose.yml and any customized configuration files, and .env files containing secrets.

Backup approach: For simple setups, stop the container, archive the volume contents, then restart. For production environments where stopping causes disruption, use filesystem snapshots or database dump commands (PostgreSQL pg_dump, SQLite .backup, MySQL mysqldump) that produce consistent backups without downtime.

For a complete automated backup workflow that ships snapshots to S3-compatible object storage, see the Restic + Rclone backup guide. Restic handles deduplication and encryption; Rclone handles multi-destination uploads. The same setup works for any Docker volume.

Backup cadence: Daily backups to remote storage are a reasonable baseline for actively used tools. Use a 30-day retention window minimum — long enough to recover from mistakes discovered weeks later. For critical data, extend to 90 days and use a secondary destination.

Restore testing: A backup that has never been restored is a backup you cannot trust. Once a month, restore your Perplexica backup to a separate Docker Compose stack on different ports and verify the data is intact. This catches silent backup failures, script errors, and volume permission issues before they matter in a real recovery.

Security Hardening

Self-hosting means you are responsible for Perplexica's security posture. The Docker Compose setup provides a functional base; production deployments need additional hardening.

Always use a reverse proxy: Never expose Perplexica's internal port directly to the internet. The docker-compose.yml binds to localhost; Caddy or Nginx provides HTTPS termination. Direct HTTP access transmits credentials in plaintext. A reverse proxy also centralizes TLS management, rate limiting, and access logging.

Strong credentials: Change default passwords immediately after first login. For secrets in docker-compose environment variables, generate random values with openssl rand -base64 32 rather than reusing existing passwords.

Firewall configuration:

ufw default deny incoming
ufw allow 22/tcp
ufw allow 80/tcp
ufw allow 443/tcp
ufw enable

Internal service ports (databases, admin panels, internal APIs) should only be reachable from localhost or the Docker network, never directly from the internet.

Network isolation: Docker Compose named networks keep Perplexica's services isolated from other containers on the same host. Database containers should not share networks with containers that don't need direct database access.

VPN access for sensitive services: For internal-only tools, restricting access to a VPN adds a strong second layer. Headscale is an open source Tailscale control server that puts your self-hosted stack behind a WireGuard mesh, eliminating public internet exposure for internal tools.

Update discipline: Subscribe to Perplexica's GitHub releases page to receive security advisory notifications. Schedule a monthly maintenance window to pull updated images. Running outdated container images is the most common cause of self-hosted service compromises.

Troubleshooting Common Issues

Container exits immediately or won't start

Check logs first — they almost always explain the failure:

docker compose logs -f perplexica

Common causes: a missing required environment variable, a port already in use, or a volume permission error. Port conflicts appear as bind: address already in use. Find the conflicting process with ss -tlpn | grep PORT and either stop it or change Perplexica's port mapping in docker-compose.yml.

Cannot reach the web interface

Work through this checklist:

Confirm the container is running: docker compose ps
Test locally on the server: curl -I http://localhost:PORT
If local access works but external doesn't, check your firewall: ufw status
If using a reverse proxy, verify it's running and the config is valid: caddy validate --config /etc/caddy/Caddyfile

Permission errors on volume mounts

Some containers run as a non-root user. If the Docker volume is owned by root, the container process cannot write to it. Find the volume's host path with docker volume inspect VOLUME_NAME, check the tool's documentation for its expected UID, and apply correct ownership:

chown -R 1000:1000 /var/lib/docker/volumes/your_volume/_data

High resource usage over time

Memory or CPU growing continuously usually indicates unconfigured log rotation, an unbound cache, or accumulated data needing pruning. Check current usage with docker stats perplexica. Add resource limits in docker-compose.yml to prevent one container from starving others. For ongoing visibility into resource trends, deploy Prometheus + Grafana or Netdata.

Data disappears after container restart

Data stored in the container's writable layer — rather than a named volume — is lost when the container is removed or recreated. This happens when the volume mount path in docker-compose.yml doesn't match where the application writes data. Verify mount paths against the tool's documentation and correct the mapping. Named volumes persist across container removal; only docker compose down -v deletes them.

Keeping Perplexica Updated

Perplexica follows a regular release cadence. Staying current matters for security patches and compatibility. The update process with Docker Compose is straightforward:

docker compose pull          # Download updated images
docker compose up -d         # Restart with new images
docker image prune -f        # Remove old image layers (optional)

Read the changelog before major version updates. Some releases include database migrations or breaking configuration changes. For major version bumps, test in a staging environment first — run a copy of the service on different ports with the same volume data to validate the migration before touching production.

Version pinning: For stability, pin to a specific image tag in docker-compose.yml instead of latest. Update deliberately after reviewing the changelog. This trades automatic patch delivery for predictable behavior — the right call for business-critical services.

Post-update verification: After updating, confirm Perplexica is functioning correctly. Most services expose a /health endpoint that returns HTTP 200 — curl it from the server or monitor it with your uptime tool.

More self-hosted AI search tools at OSSAlt.

Self-Host Perplexica: Open Source Perplexity 2026