[ChiragAgg5k]

Why

The deprecated docs MCP server at mcp-for-docs.appwrite.io is being replaced by a much smaller service that keeps its URL and tools for existing clients. Those clients have no OAuth session, so it can't use appwrite_search_docs over the MCP endpoint, and it shouldn't carry its own 35 MB index and OpenAI key.

What

• POST /docs/search returns {"results": [...]} using the existing DocsSearch.
• The route is public, like the server it supports, and rate limited per IP at the edge.
• Each call's cost is bounded: bodies are capped at 16 KiB, queries at 500 characters, and the page count is fixed at 5.
• DocsSearch.find splits the ranking from the MCP content formatting. The tool's behaviour is unchanged.

Testing

Added DocsSearchRoute to tests/e2e/test_tools.py, against the real app and index with only the OpenAI embedding stubbed: unauthenticated search, fixed page count, invalid and oversized queries, and malformed bodies. Unit suite, tests/e2e/test_tools.py, ruff and black pass locally.

[hansi-codes]

🟡 Tier B · Needs changes before merging

The earlier finding about unauthenticated access to paid embeddings remains open; this incremental change only addresses request-body buffering.

This pull request adds a public POST /docs/search endpoint for clients of the deprecated docs MCP server and separates document ranking from MCP response formatting. The route uses the existing search index, fixes the page limit, validates queries, and includes end-to-end coverage for unauthenticated search and invalid requests.

Latest changes: The route now streams incoming request bytes and returns 413 once they exceed 16 KiB, with the documentation and end-to-end test updated to cover that limit.

Verdict	New comments	Fixed	Still open
💬 Commented	0	1	1

Fix with agent prompt

### Issue 1
src/mcp_server_appwrite/http_app.py:606
**Public route leaves paid embeddings open to unlimited requests**

This now mounts an unauthenticated endpoint for every deployment with docs search enabled, and each accepted POST triggers an OpenAI embedding call. The 500-character cap only bounds an individual request; without an authentication or rate limit, anyone can repeatedly call the public route and consume the deployment's paid API quota.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

📂 Walkthrough · 6

File	Change
docs/documentation-search.md	Documents the public search route and its response and error behavior, including the 16 KiB body limit.
src/mcp_server_appwrite/constants.py	Adds the maximum query-length and request-body-size constants.
src/mcp_server_appwrite/docs_search.py	Separates document ranking from MCP content formatting.
src/mcp_server_appwrite/http_app.py	Adds the public search endpoint, fixed page count, query validation, and streaming body-size rejection.
src/mcp_server_appwrite/operator.py	Exposes the configured docs search instance to the HTTP app.
tests/e2e/test_tools.py	Covers unauthenticated search, fixed page count, invalid requests, and oversized request bodies.

⏳ Still open from earlier reviews · 1

• 🟠 Public route leaves paid embeddings open to unlimited requests · src/mcp_server_appwrite/http_app.py:606

✅ Fixed since the last review · 1

• Limit the body size before parsing JSON · src/mcp_server_appwrite/http_app.py:465

---

_{Reviewed the commits since 05170b8 · Details · Comment @hansi-codes review to re-run, or mention @hansi-codes with a question.}

[hansi-codes]

🔵 Tier A · See the inline comments. Summary

[hansi-codes]

🟢 Tier S · Looks good to merge. Summary

[hansi-codes]

🟡 Tier B · See the inline comments. Summary

[hansi-codes]

Public route leaves paid embeddings open to unlimited requests

This now mounts an unauthenticated endpoint for every deployment with docs search enabled, and each accepted POST triggers an OpenAI embedding call. The 500-character cap only bounds an individual request; without an authentication or rate limit, anyone can repeatedly call the public route and consume the deployment's paid API quota.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/mcp_server_appwrite/http_app.py
Line: 606

Comment:
**Public route leaves paid embeddings open to unlimited requests**

This now mounts an unauthenticated endpoint for every deployment with docs search enabled, and each accepted POST triggers an OpenAI embedding call. The 500-character cap only bounds an individual request; without an authentication or rate limit, anyone can repeatedly call the public route and consume the deployment's paid API quota.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{🟠 Major · security · Reply if this doesn't apply.}

[ChiragAgg5k]

Each call is bounded here, and the route is rate limited per IP at the edge before this ships.

[hansi-codes]

Per-IP edge limiting addresses repeated requests from one address, but it does not bound aggregate embedding spend: clients using many IPs can stay below each IP’s limit while continuing to make requests. Every accepted POST calls docs_search.find (route handler), which ranks the query (embedding path). So the per-call bounds and per-IP limit still leave aggregate paid usage unbounded; a global quota or equivalent cap would close that gap.

[ChiragAgg5k]

This does not apply any more: a per-IP rate limit on this path is now live at the edge in front of the hosted server, at 60 requests a minute. It sits outside this repository, so it does not appear in this diff. Together with the 16 KiB body cap, the 500-character query cap and the fixed page count, that bounds what the route can spend. Keeping the route unauthenticated is deliberate, because the deprecated server it supports is also public.

[hansi-codes]

I accept that the edge rule is live. A limit of 60 requests per IP per minute still does not bound aggregate spend: requests from many distinct IPs can each stay within the limit, and each accepted request reaches the embedding call (route, embedding). The request-size and query limits bound individual calls, not their total. The finding remains open unless there is also an aggregate cap or equivalent control.

[hansi-codes]

🟡 Tier B · See the inline comments. Summary