Skip to content

DEPLOY.md

Latest commit

 

History

History
104 lines (77 loc) · 4.12 KB

DEPLOY.md

File metadata and controls

104 lines (77 loc) · 4.12 KB
Raw

Deploying to Vercel + Railway + Qdrant Cloud

This repo is set up for a three-service split:

Piece Where Cost
Frontend (React/Vite static build) Vercel Free
Backend (FastAPI + UniXcoder) Railway ~$5–10/mo
Vector database Qdrant Cloud Free tier (1 GB)

Total setup time: ~30 minutes plus the migration copy (a few minutes on top).

1. Create a Qdrant Cloud cluster

  1. Sign up at https://cloud.qdrant.io.
  2. Create a free-tier cluster (1 GB is plenty for this dataset).
  3. Copy the cluster URL and generate an API key. You will need both below.

2. Migrate the collections into Qdrant Cloud

Make sure your local Qdrant container is running and has these collections:

  • code-files (file contents used by the file endpoint)
  • code-signatures (MiniLM embeddings of function signatures)
  • code-snippets-unixcoder (UniXcoder embeddings; still building on first run)

Then copy them into the cloud cluster:

export SRC_URL=http://localhost:6333
export DST_URL=https://your-cluster-id.aws.cloud.qdrant.io:6333
export DST_API_KEY=your-cloud-api-key
python -m tools.migrate_to_qdrant_cloud

If a collection is not yet ready locally (for example the UniXcoder run is still going), the script prints skip for it and moves on. Re-run once it's ready.

3. Deploy the backend to Railway

  1. Push this repo to GitHub.

  2. In Railway, click New Project → Deploy from GitHub Repo and pick this repo. Railway detects Dockerfile and railway.json.

  3. Set the following service variables in Railway:

    Variable Value
    QDRANT_URL your Qdrant Cloud URL (e.g. https://xxx.aws.cloud.qdrant.io:6333)
    QDRANT_API_KEY your Qdrant Cloud API key
    CORS_ORIGINS your Vercel frontend URL (see step 4). Comma-separate multiple.
    WORKERS 1 (raise only if you upgrade RAM significantly)
    INDEXED_COMMIT the qdrant/qdrant SHA the collections were built from, printed at the end of the indexing log. Result links carry line numbers and only match at that commit; leave it unset and they resolve against master, drifting as the source moves.

  4. Railway will build and expose a URL like https://code-search-api.up.railway.app. Wait for the health check at /api/health to pass; first boot takes ~30–60s because the UniXcoder model loads into memory.

  5. Copy that URL — you'll paste it into Vercel next.

Sizing note. UniXcoder plus MiniLM want about 2 GB of RAM to be comfortable. The free/Hobby "Starter" instance (512 MB) will OOM. Use Railway's Hobby plan (2 GB) or higher.

4. Deploy the frontend to Vercel

  1. In Vercel, click Add New → Project and import this repo.

  2. Set the Root Directory to frontend. Vercel picks up vercel.json automatically.

  3. Under Environment Variables, add:

    Variable Value
    VITE_API_URL your Railway backend URL, no trailing slash

  4. Click Deploy. Vercel gives you a URL like https://code-search.vercel.app.

  5. Go back to Railway and update CORS_ORIGINS with that exact URL. Railway redeploys automatically.

5. Verify

  • Hit the frontend URL — you should see the hero.
  • Run one of the demo queries. You should see results (semantic if UniXcoder is loaded, "Warming Up" keyword mode otherwise).
  • Click into a code card and hit "load more lines" — the file endpoint should return 200s.

Troubleshooting

  • Search returns 500 with a Qdrant 404 error: the target collection was not migrated. Re-run python -m tools.migrate_to_qdrant_cloud.
  • CORS error in browser console: the frontend URL isn't in CORS_ORIGINS on Railway. Add it (comma-separated for multiple).
  • Backend crashes on boot with OOM: raise the Railway plan to at least 2 GB, or switch to a MiniLM-only build (drop UniXcoder from CodeSearcher).
  • Slow first request: normal. UniXcoder loads once per container, then requests are fast.

Custom domain

Both Vercel and Railway support custom domains. If you attach one to the frontend, remember to update CORS_ORIGINS on Railway to include it.