Skip to content

readme: docker-first quickstart, fix dead links, typography cleanup#176

Open
fm1320 wants to merge 10 commits into
mainfrom
readme-docker-first
Open

readme: docker-first quickstart, fix dead links, typography cleanup#176
fm1320 wants to merge 10 commits into
mainfrom
readme-docker-first

Conversation

@fm1320
Copy link
Copy Markdown
Collaborator

@fm1320 fm1320 commented May 18, 2026

Summary

  • Quickstart now leads with docker run and drops the pip install sie-server path. Docker is the canonical local path; the pip route added env friction without paying for itself.
  • SDK example is unchanged (still targets http://localhost:8080, which is what the docker command exposes).
  • Fix npm package name in the TypeScript line: @sie/sdk -> @superlinked/sie-sdk (the published name in packages/sie_ts_sdk/package.json). The old command 404s.
  • Repoint all docs links from sie.dev/docs/... to superlinked.com/docs/.... sie.dev is now a 308 redirect shim; use the canonical domain.
  • Models reference link: sie.dev/docs/reference/models/ was already 404 (pre-existing dead link). Replaced with https://superlinked.com/models, the actual 85+ model catalog page.
  • Footer text and href updated to superlinked.com/docs.
  • Typography pass: remove em dashes, middle dots, and the decorative -> arrows on link text. Nav and footer use |, prose lists use commas, link-pair lines use periods.

Verification

  • All 18 external URLs in the final README return 200 with redirects followed (curl).
  • SDK example signatures match packages/sie_sdk/src/sie_sdk/client/sync.py:
    • encode(model, Item) returns EncodeResult["dense"]
    • score(model, query, [items]) returns ScoreResult["scores"]
    • extract(model, Item, labels=[...]) returns ExtractResult["entities"]
  • Docker tags latest-cpu-default and latest-cuda12-default confirmed live on GHCR.

Test plan

  • Render check on GitHub for layout
  • Pull the CPU image and confirm docker run -p 8080:8080 ghcr.io/superlinked/sie-server:latest-cpu-default boots
  • Run the SDK snippet end-to-end against the running container

fm1320 added 2 commits May 18, 2026 12:47
- Quickstart leads with docker run; remove pip install sie-server
- Fix npm package name to @superlinked/sie-sdk
- Repoint docs links from sie.dev to superlinked.com
- Replace dead reference/models page with superlinked.com/models
- Footer now superlinked.com/docs
- Remove em dashes from Explore section
@fm1320 fm1320 changed the title readme: docker-first quickstart, fix dead links, drop em dashes readme: docker-first quickstart, fix dead links, typography cleanup May 18, 2026
fm1320 added 8 commits May 18, 2026 12:53
- One-sentence framing of the server+SDK architecture at top of Quickstart
- Steps renamed: "Run the engine" / "Call it from Python"
- Note image sizes (~1 GB CPU, ~9 GB GPU) and first-model-load time
- Note SDK is a thin HTTP client, not a model runtime
- Remove the Colab badge line above step 1 (notebook still linked in Explore)
- Reframe quickstart: explicitly tell readers to start docker first, then SDK
- Rename step 2 from 'Call it from Python' to language-neutral 'Call it'
- Mount a named docker volume for the HF cache so model weights persist
  across runs; chose named volume over host bind-mount because the container
  runs as non-root sie user and bind-mounts hit UID issues on macOS/Windows
- Replace the 'ready in seconds' claim with a measured statement: container
  reaches /readyz in a few seconds; verified at 5.3s on Apple Silicon under
  Rosetta against ghcr.io/superlinked/sie-server:latest-cpu-default
- Verified end to end: encode(Stella 400M, Hello world) -> shape (1024,);
  warm call returns in ~0.1s
- Note --platform linux/amd64 requirement for Apple Silicon
- Step 2 was 'Call it', too generic; rename to call out both supported SDKs
- Show pip and pnpm install commands side by side in step 2
- Trim the redundant TypeScript line below the Python example to a pointer at the TS docs
…code

The dense post-docker paragraph was a wall of caveats that buried the hook.
Keep only the one-liner that unblocks pull on Apple Silicon and the link to the
deployment guide. Move the 'first model call is slow' note under the code
example, where a dev would actually hit it.
The 'Apple Silicon: add --platform linux/amd64' footnote was easy to miss;
most local dev happens on Mac, so the macOS line is now first-class and the
--platform flag is baked into the command itself. Verified all three lines:
- macOS (Apple Silicon): /readyz 200 after 6s under Rosetta
- Linux CPU: same image, same command without --platform
- Linux GPU: --gpus all on the cuda12 image

Footer now links to the Superlinked talk at the AI Engineer conference.
The Rosetta/platform explanation was for chat, not docs. The flag is in
the command line; that is enough for a dev trying things out quickly.
Footer link reframed from 'Watch the AI Engineer talk' to 'Why we built
SIE (AI Engineer Europe 2026)' so the reason to click is the motivation
story, not the format.
1. Cut the framing paragraph from three sentences to one
2. Move 'first call is slow' note from below the code into the code itself,
   right where a dev will be staring at the call
3. Swap the example to small models so first run is ~40s instead of ~10 min:
   - encode:  all-MiniLM-L6-v2 (~90 MB, 384-dim)
   - score:   ms-marco MiniLM L-6 (~80 MB)
   - extract: GLiNER multi v2.1 (unchanged)
   Verified live: encode -> shape (384,), elapsed 40s on Apple Silicon under
   Rosetta; warm calls in ms.
4. Add 'curl /readyz' sanity-check between Step 1 and Step 2
5. Drop the redundant 'three functions' bullet from About
6. Helm placeholder <TOKEN> -> YOUR_HF_TOKEN
7. Bridge from Explore catalog display names to HF IDs used in the SDK
… outputs

- Move 'Why we built SIE' from footer up into the Explore section, alongside
  Notebooks and Examples. Footer is a footnote and a CTA does not belong there.
- Drop the 'See the deployment guide for GPU pinning, read-only filesystems,
  and tuning' line. None of those are local-dev concerns; the deployment
  guide is still linked from the Production section where it belongs.
- Fix example output values to match what the models actually return,
  verified end-to-end against ghcr.io/superlinked/sie-server:latest-cpu-default:
    * encode(all-MiniLM-L6-v2): shape (384,)
    * score(ms-marco-MiniLM-L-6-v2): logits -7.1 / -11.048 (not 0.998/0.012;
      cross-encoder outputs raw logits, not probabilities)
    * extract(gliner_multi-v2.1): Tim Cook=0.991, Apple=0.978
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant