Die Oberfläche ist lokalisiert; Skill-Beschreibungen und -Dokumente werden per KI automatisch in Ihre Sprache übersetzt. Code, Befehle, Pfade und Fachbegriffe bleiben im Original, maßgeblich ist das Dokument des Autors.
Skills entdecken
Daten und Forschung · SKILL.md

integrations

integrations

Work on LangChain partner integration packages under libs/partners and provider model-profile data: package metadata, provider clients, chat/LLM/embedding/vector store classes, credentials, tests, cassettes, optional dependencies, and profile refresh workflows.

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill-Dokumentation

GitHub

name: integrations description: "Work on LangChain partner integration packages under libs/partners and provider model-profile data: package metadata, provider clients, chat/LLM/embedding/vector store classes, credentials, tests, cassettes, optional dependencies, and profile refresh workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

Integrations

Use this sub-skill for practical work on LangChain-maintained partner packages and provider model-profile data. It covers provider-specific packages such as OpenAI, Anthropic, Ollama, Qdrant, Chroma, DeepSeek, Fireworks, Groq, Hugging Face, MistralAI, Nomic, Perplexity, OpenRouter, xAI, and Exa.

When to Use

  • The user asks to add, change, or debug a provider package under libs/partners/<provider>.
  • The work touches provider chat models, LLMs, embeddings, vector stores, retrievers, tools, provider clients, middleware, output parsers, optional dependencies, credentials, or environment-variable handling.
  • The user asks about package-local pyproject.toml, extras/dependency groups, import/version scripts, provider-specific unit or integration tests, VCR/cassette isolation, or model-profile refresh for provider data directories.
  • The user needs safe validation that avoids network calls by default and escalates only when credentials, services, and permission are available.

Route Elsewhere

  • For package-agnostic langchain-core primitives, use ../core-primitives/SKILL.md.
  • For high-level langchain v1 agent construction, middleware routing, or init_chat_model behavior outside a provider package, use ../agents-and-middleware/SKILL.md.
  • For monorepo-wide package metadata, CI labels, release wiring, contributor conventions, or root development workflow, use ../monorepo-development/SKILL.md.
  • For implementing shared standard-test suites or model-profile CLI internals, route to ../testing-and-profiles/SKILL.md when that sibling is present; this sub-skill only covers partner-side usage and provider data refresh.

Reference Map

  • Start with references/provider-patterns.md for partner package layout, representative public classes, dependency patterns, source/test locations, edit workflows, and validation signals.
  • Use references/credentials-and-testing.md for API keys, local service checks, no-network unit tests, integration-test skip conditions, VCR cassette handling, and OpenAI Codex cassette safety.
  • Use references/model-profiles.md for provider profile data directories, langchain-profiles refresh, augmentation files, generated profile diffs, and profile drift review.
  • Use references/troubleshooting.md for missing optional SDKs, credential failures, cassette isolation, model-profile drift, vector store service availability, and provider response/tool-call shape mismatches.
  • Run scripts/partner_package_check.py for a read-only metadata/layout check of a partner package when a Python interpreter is available.

Fast Workflow

  1. Identify the provider package, then work from its package directory: libs/partners/<provider>.
  2. Inspect pyproject.toml, the langchain_<provider>/ package root, tests/unit_tests/, tests/integration_tests/, and package-local scripts/ before editing.
  3. Keep provider-specific behavior in the partner package. Shared interfaces, message schemas, callbacks, base vector-store contracts, and standard-test framework changes belong to sibling skills.
  4. Add deterministic unit tests for new behavior. Network-backed tests must be opt-in, credential-gated, and isolated from default no-network unit test runs.
  5. Validate narrowly first with package-local uv commands when uv is available; do not use pip, poetry, conda, or ad-hoc virtual environments for this monorepo.

Safe Validation

From a partner package directory, prefer targeted checks such as:

uv run --group test pytest tests/unit_tests/test_imports.py
uv run --group test pytest tests/unit_tests/test_chat_models.py
uv run --group test pytest tests/unit_tests/test_embeddings.py
uv run --group test pytest tests/unit_tests/test_vectorstores.py
uv run --group lint python scripts/check_imports.py $(find langchain_* -name '*.py')
uv run python scripts/check_version.py

Adjust test filenames to the package. Some packages use make test, make check_imports, make check_version, make lint_package, or make type; run these from the package directory only after confirming the package Makefile supports them. Skip network-heavy integration tests unless the user explicitly provides credentials or a local service and asks for those checks.

For a safe metadata/layout check that does not import provider SDKs or contact services:

python skills/langchain/sub-skills/integrations/scripts/partner_package_check.py libs/partners/openai

Guardrails

  • Preserve public provider class signatures and exported names unless the user explicitly accepts a breaking change.
  • Do not add a provider SDK dependency to an unrelated package; partner SDKs belong in that partner package's metadata.
  • Do not run credentialed examples, cassette recording, vector database services, or live provider calls without explicit user permission.
  • Do not manually edit generated _profiles.py as the primary fix for profile drift; refresh through the model-profile workflow and review the generated diff.
  • Do not link runtime instructions to source checkout docs, examples, or scripts; use the bundled references and scripts in this sub-skill.
Gespeicherte Skills