qdrant-loader-core 0.7.4

Shared core for provider-agnostic LLM support and configuration mapping for qdrant-loader ecosystem

QDrant Loader Core

Shared core library for the QDrant Loader ecosystem. It provides a provider‑agnostic LLM layer (embeddings and chat), configuration mapping, safe logging, and normalized error handling used by the CLI and MCP Server packages.

🚀 What It Provides

• Provider‑agnostic LLM facade: OpenAI, Azure OpenAI, OpenAI‑compatible, and Ollama
• Unified async APIs:
  • EmbeddingsClient.embed(inputs: list[str]) -> list[list[float]]
  • ChatClient.chat(messages: list[dict], **kwargs) -> dict
  • TokenCounter.count(text: str) -> int
• Typed settings and mapping: LLMSettings.from_global_config(...) supports the new global.llm schema and maps legacy fields with deprecation warnings
• Structured logging with redaction: LoggingConfig.setup(...) masks secrets and reduces noisy logs
• Normalized errors: consistent exceptions across providers (TimeoutError, RateLimitedError, InvalidRequestError, AuthError, ServerError)
• Optional dependencies via extras: openai, ollama

📦 Installation

# Minimal core
pip install qdrant-loader-core

# With OpenAI/Azure OpenAI support
pip install "qdrant-loader-core[openai]"

# With Ollama support
pip install "qdrant-loader-core[ollama]"

# From source (development)
git clone https://github.com/martin-papy/qdrant-loader.git
cd qdrant-loader
pip install -e packages/qdrant-loader-core

⚡ Quick Start

Example using the new global.llm schema:

global:
  llm:
    provider: "openai"            # openai | azure_openai | ollama | openai_compat
    base_url: "https://api.openai.com/v1"
    api_key: "${LLM_API_KEY}"
    models:
      embeddings: "text-embedding-3-small"
      chat: "gpt-4o-mini"

import asyncio
from qdrant_loader_core.llm.settings import LLMSettings
from qdrant_loader_core.llm.factory import create_provider

global_config = {
    "llm": {
        "provider": "openai",
        "base_url": "https://api.openai.com/v1",
        "api_key": "${LLM_API_KEY}",
        "models": {"embeddings": "text-embedding-3-small", "chat": "gpt-4o-mini"},
    }
}

settings = LLMSettings.from_global_config(global_config)
provider = create_provider(settings)

async def main() -> None:
    vectors = await provider.embeddings().embed(["hello", "world"])  # list[list[float]]
    reply = await provider.chat().chat([
        {"role": "system", "content": "You are helpful."},
        {"role": "user", "content": "Say hi!"},
    ])
    print(len(vectors), reply["text"])  # 2 "Hi!" (example)

asyncio.run(main())

🔌 Supported Providers

• OpenAI ([openai] extra): Uses the official openai Python SDK. Configure with base_url, api_key, and models.chat/models.embeddings.
• Azure OpenAI ([openai] extra): Requires api_version. Auto‑detected when the host is *.openai.azure.com or *.cognitiveservices.azure.com. Optional provider_options.azure_endpoint overrides the endpoint.
• OpenAI‑compatible ([openai] extra): Any endpoint exposing OpenAI‑style /v1 APIs. Set provider: openai_compat (or rely on base_url containing openai).
• Ollama ([ollama] extra): Works with native /api and OpenAI‑compatible /v1 endpoints. Optional provider_options.native_endpoint: auto | embed | embeddings selects native behavior.

🔧 Configuration Mapping

LLMSettings.from_global_config(...) accepts a parsed dict for global and supports:

• New schema (recommended): global.llm

  • provider, base_url, api_key, api_version (Azure), headers
  • models: { embeddings, chat }
  • tokenizer
  • request: { timeout_s, max_retries, backoff_s_min, backoff_s_max }
  • rate_limits: { rpm, tpm, concurrency }
  • embeddings: { vector_size }
  • provider_options: provider‑specific opts (e.g., azure_endpoint, native_endpoint)

• Legacy mapping (deprecated): global.embedding.* and file_conversion.markitdown.llm_model

  • Maps to provider + models (embeddings/chat), emits a deprecation warning
  • Prefer migrating to global.llm for clarity and future features

🧾 Logging

Use the built‑in structured logging with redaction:

from qdrant_loader_core.logging import LoggingConfig

LoggingConfig.setup(level="INFO", format="console", file=None)
logger = LoggingConfig.get_logger(__name__)
logger.info("LLM ready", provider=settings.provider)

Notes

• Secrets (keys/tokens) are masked in both stdlib and structlog output
• Noisy third‑party logs are toned down; Qdrant version checks are filtered
• For MCP integration, set MCP_DISABLE_CONSOLE_LOGGING=true to disable console output

🧰 Error Handling

Catch provider‑normalized exceptions from qdrant_loader_core.llm.errors:

• TimeoutError — request timed out
• RateLimitedError — rate limit exceeded
• InvalidRequestError — bad parameters or unsupported operation
• AuthError — authentication/authorization failed
• ServerError — transport/server failures

📚 Documentation

• Website — Project site and guides
• Core package docs — Package‑specific page
• Monorepo docs — Source documentation in this repository

🤝 Contributing

This package is part of the QDrant Loader monorepo. See the main contributing guide.

🆘 Support

• Issues: https://github.com/martin-papy/qdrant-loader/issues
• Discussions: https://github.com/martin-papy/qdrant-loader/discussions

📄 License

Licensed under the GNU GPLv3 — see LICENSE.