Two complementary changes to mitigate intermittent TLS handshake
timeouts and "notifications/cancelled: Bad Request" seen against the
single-worker uvicorn deployment.
1. bedesten rate-limiter back-pressure
- Add optional ``max_wait`` to ``_TokenBucket.acquire``: if the next
wait would exceed it, raise ``BedestenRateLimited`` immediately
instead of sleeping. After a server-side 429 the bucket pauses for
up to 30s; previously a queued request sat in ``asyncio.sleep``
for that whole window, holding the worker slot and pushing the
MCP client past its cancellation timeout.
- ``search_bedesten_unified`` / ``get_bedesten_document_markdown``
catch ``BedestenRateLimited`` and reuse the existing structured
429-style response, so callers get a fast, clean retry signal.
- Tunable via ``BEDESTEN_RATE_MAX_WAIT_S`` (default 8.0s).
2. Offload sync markitdown conversions to a thread
- Every ``markitdown.convert*`` call site is now wrapped in
``asyncio.to_thread(...)`` across 14 modules (bedesten, yargitay,
danistay, anayasa norm + bireysel, uyusmazlik, emsal, rekabet,
gib, kvkk, sayistay, bddk, sigorta_tahkim, kik_v2). PDF / large
HTML parsing was stalling the event loop for seconds, which on a
single-worker deployment delayed every other in-flight request
and queued new TLS handshakes until they timed out.
Verified locally:
- ``ast.parse`` + ``importlib.import_module`` on all 15 modified files
- ``mcp_server_main.create_app()`` constructs successfully
- New ``_TokenBucket.acquire(max_wait=...)`` smoke-tested across 6
paths: capacity-available, no-arg backward compat, max_wait raise,
max_wait wait+succeed, ``penalize_until`` + max_wait fast-raise.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The fetch tool's metadata lookup constructed BedestenSearchData without
the required itemTypeList field, causing a Pydantic validation error and
losing the chance to enrich the response with a proper title.
When the Bedesten API rate-limits us, point the model at the local
yargi-cli tool (https://github.com/saidsurucu/yargi-cli) so the user
has a working alternative while waiting out the limit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Bedesten API can intermittently return HTTP 429 Too Many Requests.
Previously the tool raised, leaving the LLM with an unhandled error.
Now search_bedesten_unified returns a dict with error="rate_limit_exceeded"
and get_bedesten_document_markdown returns a BedestenDocumentMarkdown
whose markdown_content describes the rate limit, so the model can
inform the user and retry. Non-429 errors still propagate.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a LocalEmbedder that targets any OpenAI-compatible embedding
endpoint (Ollama, llama.cpp, vLLM, LM Studio, ...). Zero new
Python dependencies — reuses the existing openai SDK with a
custom base_url. Defaults to Ollama at http://localhost:11434/v1
with nomic-embed-text @ 768 dims; override via env vars for other
servers/models (e.g. bge-m3 @ 1024 dims for better Turkish).
Refactors the shared encode/similarity logic into a private base
class so OpenRouterEmbedder and LocalEmbedder don't duplicate ~50
lines. OpenRouter keeps its ranking headers; local sends none.
Adds get_embedder() factory selecting the provider based on
EMBEDDING_PROVIDER (local) or OPENROUTER_API_KEY presence, and
is_semantic_search_available() that returns True for either path.
mcp_server_main now uses these so the semantic_search tool is
exposed when only a local server is configured.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
google/gemini-embedding-001 became paid on OpenRouter, leaving
users without credit unable to run the semantic_search tool. The
old code hardcoded the model and 3072 dimensions in three places.
Make OpenRouterEmbedder accept model/dimension via constructor
args or OPENROUTER_EMBEDDING_MODEL / OPENROUTER_EMBEDDING_DIMENSION
env vars, with the previous values as backward-compatible defaults.
Switch the VectorStore and the response payload in mcp_server_main
to read embedder.dimension instead of the hardcoded 3072 so a
configured non-Gemini model does not produce shape mismatches.
Bad dimension input (non-int or non-positive) now raises a clear
ValueError instead of a downstream shape error.
Documented the new env vars in .env.example.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Introduces two tools backed by the gib.gov.tr public JSON API
(reverse-engineered from the Next.js SPA chunks):
- search_gib_ozelge: keyword, ozelgeNo, kanunNo, date-range, paging
over 18k+ Revenue Administration tax rulings. Simple YYYY-MM-DD
dates are auto-expanded to ISO 8601 to satisfy the backend.
- get_gib_ozelge_document_markdown: fetch a single ruling by numeric
id and return 5000-char paginated Markdown with a metadata header
block (title, ozelgeNo, tarih, kanun, kaynak).
Also prunes stale auth/Fly.io-era entries from uv.lock.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add Insurance Arbitration Commission integration with Tavily search
and direct PDF download for 64 quarterly journal issues (2010-2025).
Tools:
- search_sigorta_tahkim_decisions: Search via Tavily API
- get_sigorta_tahkim_document_markdown: PDF download + paginated markdown
- search_within_sigorta_tahkim_issue: Keyword search within individual
decisions of a journal issue, with Turkish İ/I case folding support
Total tools: 25 (was 22)
Monkey-patch JSONRPCNotification to use extra="forbid" so that
requests with "id": null are no longer misclassified as notifications
(202 Accepted). They now correctly fail validation and return a
-32600 Invalid Request error.
- Replace EmbeddingGemma local model with OpenRouter API integration
- Use google/gemini-embedding-001 model via OpenRouter (3072 dimensions)
- Add conditional tool registration: auto-disable if OPENROUTER_API_KEY not set
- Add openai and numpy dependencies to pyproject.toml
- Update .env.example with OPENROUTER_API_KEY configuration
- Fix ruff lint issues in semantic_search module
- Add global _health_check_client variable for singleton pattern
- Define get_or_create_health_check_client() function for health checks
- Add cleanup for health check client in perform_cleanup()
Fixes Bedesten health check error: "name 'get_or_create_health_check_client' is not defined"
- Fixed 10 tools to avoid / patterns in JSON schemas
- All tools now return Dict[str, Any] with .model_dump() applied
- Affected tools:
* search_emsal_detailed_decisions
* get_emsal_document_markdown
* search_uyusmazlik_decisions
* get_uyusmazlik_document_markdown_from_url
* search_rekabet_kurumu_decisions
* get_rekabet_kurumu_document
* search_sayistay_unified
* get_sayistay_document_unified
* search_kvkk_decisions
* get_kvkk_document_markdown
- Gemini CLI should now be able to load and use all MCP tools without schema validation errors
Resolves 'cannot convert undefined or null to object' error in search_bedesten_unified
by adding proper null checking for response.data.emsalKararList and response.data.total
fields before accessing them.
- Add hasattr() and null checks for response.data fields
- Provide safe defaults: empty list for emsalKararList, 0 for total
- Prevents TypeError when API returns undefined/null fields
- Matches null safety pattern used in other search tools
- Disable issuer validation in BearerAuthProvider (issuer=None)
- Simplify authentication condition (remove auth_enabled check)
- Revert CORS middleware to simple configuration
- Fix OAuth metadata endpoint to match v0.1.6
- Apply conditional auth only to MCP server creation
Critical fixes for Claude AI tools discovery
- Remove Redis session store initialization from create_app()
- Revert to simple token counting middleware only
- Fix session management issue causing tools to appear then disappear
- This matches the exact v0.1.6 implementation that was working
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Revert asgi_app.py to v0.1.6 approach with path='/' for MCP app
- Fix uyusmazlik client: use individual httpx.AsyncClient instead of shared
- Fix health check: use individual httpx.AsyncClient instead of shared
- Remove shared_health_check_client that was causing connection drops
