langchaint
langchaint provides typed generation and embeddings with raw provider responses and per-attempt billing. Alpha: the API is unstable and may change without notice.
Purpose
langchaint provides BoundLLM and EmbeddingModel as provider-neutral async interfaces.
langchaint provides no agent class or agent loop.
langchaint uses pyrefly with the strict all preset.
Applications should use langchaint with a strict type checker.
langchaint relies on static checking for argument types.
langchaint does not repeat every argument type check at runtime.
Install and authenticate
langchaint requires Python 3.13 or newer.
Applications install and pin each provider SDK directly.
langchaint declares no dependency extras.
Top-level import langchaint requires no provider SDK or numpy.
Backend imports report missing dependencies through ModuleNotFoundError.
| Provider | Class | Creates | Depends | Default credentials |
|---|---|---|---|---|
| Anthropic | Anthropic |
LLM |
anthropic |
ANTHROPIC_API_KEY |
| Amazon Bedrock | AnthropicBedrock |
LLM |
anthropic[bedrock] |
AWS credential provider chain |
| Amazon Bedrock | CohereBedrock |
EmbeddingModel |
boto3 and numpy |
AWS credential provider chain |
| DeepSeek | DeepSeek |
LLM |
openai |
DEEPSEEK_API_KEY |
| Gemini | Gemini |
LLM |
google-genai |
GOOGLE_API_KEY or GEMINI_API_KEY |
| OpenAI | OpenAI |
LLM, EmbeddingModel |
openai (numpy and tiktoken for embeddings) |
OPENAI_API_KEY |
| Amazon Bedrock | OpenAIBedrock |
LLM |
openai[bedrock] |
AWS credential provider chain |
Every listed class accepts client= for SDK client configuration.
The Amazon Bedrock classes use the AWS credential provider chain.
This includes environment credentials, profiles, SSO, containers, and instance roles.
Pass aws_region= to select a Bedrock region explicitly.
Generation quickstart
import asyncio
from langchaint.openai import OpenAI
async def main() -> None:
openai = OpenAI()
bound_llm = openai.model("gpt-5.6-terra", regional_processing=False).bind(
system_prompt="Answer clearly and concisely.",
automatic_prompt_caching=False,
)
response = await bound_llm.generate_one("Why is the sky blue?")
print(response.output, response.usage.cost_in_usd)
asyncio.run(main())
response.output is assistant text.
regional_processing states whether the configured endpoint uses regional processing.
Pass a Pydantic model to LLM.bind(response_format=...) for validated structured output.
Embedding quickstart
import asyncio
from langchaint.openai import OpenAI
async def main() -> None:
openai = OpenAI()
embedding_model = openai.embedding_model(
"text-embedding-3-small",
dimension=1024,
)
documents = await embedding_model.embed(
["The Moon orbits Earth.", "Mars has two small moons."],
task="retrieval_document",
)
query = await embedding_model.embed(
["Which object circles Earth?"],
task="retrieval_query",
)
print(documents.shape, query.shape)
asyncio.run(main())
EmbeddingModel.embed() requires task for every adapter.
The OpenAI adapter sends no corresponding request field.
EmbeddingModel.embed() returns normalized Float2D values with numpy.float32 elements.
Each input produces one row.
Use CohereBedrock for Cohere embeddings through Amazon Bedrock.
Share one rate-limit quota
Create one OpenAI per rate-limit quota.
openai = OpenAI(
max_concurrent_requests=8,
max_request_starts_per_second=50.0,
)
terra = openai.model("gpt-5.6-terra", regional_processing=False)
sol = openai.model("gpt-5.6-sol", regional_processing=False)
Both terra and sol use openai.client and one SharedBackoff.
max_concurrent_requests applies across both LLM values.
max_request_starts_per_second applies across both LLM values.
An EmbeddingModel from openai.embedding_model() uses the same client and SharedBackoff.
Pass an SDK client to close it directly.
from openai import AsyncOpenAI
client = AsyncOpenAI()
openai = OpenAI(client=client)
terra = openai.model("gpt-5.6-terra", regional_processing=False)
# Use terra.
await client.close()
Binding and results
Call LLM.bind() before generating.
BoundLLM provides generate_one, generate_many, and stream_one.
BoundLLM.rebind() replaces selected binding fields.
Pass tools=[tool] to LLM.bind(), then dispatch through BoundLLM.tool_manager.
LLM.bind(max_attempts=...) limits requests for one GenerationInput, including the first.
An embedding batch contains inputs sent together during each attempt.
embedding_model(max_attempts=...) limits requests for one embedding batch, including the first.
automatic_prompt_caching is required because it changes billing.
cache_breakpoint=True ends the reusable prefix at that ContentPart.
GenerateResult and GenerationError include paid Usage across attempts.
More examples
examples/README.md indexes focused examples and migration guidance.
The examples cover structured output, batches, streaming, tools, tracing, pricing, and failures.
They also cover prompt caching, reasoning, embeddings, and complete application structure.
Development
Run scripts/CI.sh before committing.
The tests are offline and require no API keys.
License
langchaint uses the MIT License.
1 maintainer
Release files for langchaint 0.16.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| langchaint-0.16.0.tar.gz | 219.6 kB | Details |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| langchaint-0.16.0-py3-none-any.whl | Python 3 | none | any | Details |
Total release size: 457.0 kB
Release files / langchaint-0.16.0.tar.gz
| Download URL | langchaint-0.16.0.tar.gz |
|---|---|
| Size | 219.6 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
ab62a16249603dd14dfd38c2e94461105fcfca98f22faaa16b6900819570adeb
|
|
BLAKE2b-256 checksum How to use checksums |
d903c9b675f1666432d081c94b7a429b04d5f14bc2b84adb993d7c1494ef671c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
uv/0.12.3 {"installer":{"name":"uv","version":"0.12.3","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}
|
Release files / langchaint-0.16.0-py3-none-any.whl
| Download URL | langchaint-0.16.0-py3-none-any.whl |
|---|---|
| Size | 237.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
1646732dbd2ab1934ded23c0e7b59c5ef858916aad42d4d86b1db10c3e67ca50
|
|
BLAKE2b-256 checksum How to use checksums |
ca9f68c51c06fc30401f265f9cf706ee96836d587a5860a9907e571a2766404d
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
uv/0.12.3 {"installer":{"name":"uv","version":"0.12.3","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"24.04","id":"noble","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}
|
