Skip to main content

Run AI tasks and workflows locally.

Project description

AIBackends

Run AI tasks and workflows locally.

Build extraction, classification, embeddings, redaction, and analysis pipelines in plain Python with llamacpp and transformers.

  • First-class llamacpp and transformers runtimes
  • Typed outputs for extraction and analysis tasks
  • Reusable tasks and workflows for scripts, apps, and batch jobs
  • Practical local examples for text, image OCR, documents, audio, and video

Install

pip install aibackends

# Local runtimes
pip install aibackends[llamacpp]
pip install aibackends[llamacpp-cuda]
pip install aibackends[llamacpp-metal]
pip install aibackends[transformers]

# Capability extras
pip install aibackends[pdf]
pip install aibackends[audio]
pip install aibackends[video]
pip install aibackends[pii]

Quickstart

Extract an invoice locally

from aibackends.models import GEMMA4_E2B
from aibackends.runtimes import LLAMACPP
from aibackends.tasks import ExtractInvoiceTask, create_task

task = create_task(
    ExtractInvoiceTask,
    runtime=LLAMACPP,
    model=GEMMA4_E2B,
)

result = task.run("invoice.pdf")
print(result.total)

Examples

Single tasks

Classify text locally and redact PII

from aibackends.models import GEMMA4_E2B
from aibackends.runtimes import LLAMACPP
from aibackends.tasks import ClassifyTask, RedactPIITask, create_task

classifier = create_task(
    ClassifyTask,
    runtime=LLAMACPP,
    model=GEMMA4_E2B,
    labels=["invoice", "contract", "receipt"],
)
redactor = create_task(
    RedactPIITask,
    backend="gliner",
    labels=["email", "phone_number"],
)

classification = classifier.run("invoice text")
redacted = redactor.run("john@example.com called from +1 555 0100")

RedactPIITask uses a dedicated backend such as gliner or openai-privacy (the local privacy-filter model) rather than the general LLM runtime interface.

Generate local embeddings

from aibackends.models import MINILM_L6
from aibackends.runtimes import TRANSFORMERS
from aibackends.tasks import EmbedTask, create_task

embedder = create_task(
    EmbedTask,
    runtime=TRANSFORMERS,
    model=MINILM_L6,
)

vector = embedder.run("Payments failed after checkout deploy.")
print(len(vector))
print(vector[:5])

Workflows

Batch-process sales calls analysis locally

from pathlib import Path

from aibackends.models import GEMMA4_E2B
from aibackends.runtimes import LLAMACPP
from aibackends.workflows import SalesCallAnalyser, create_workflow

workflow = create_workflow(
    SalesCallAnalyser,
    runtime=LLAMACPP,
    model=GEMMA4_E2B,
)

results = workflow.run_batch(
    inputs=Path("./calls").glob("*.m4a"),
    max_concurrency=4,
    on_error="collect",
)

Run OCR locally

from pydantic import BaseModel, Field

from aibackends.models import QWEN3_VL_4B
from aibackends.runtimes import LLAMACPP
from aibackends.schemas.common import LineItem
from aibackends.steps.enrich import VisionExtractor
from aibackends.steps.ingest import ImageIngestor
from aibackends.workflows import Pipeline


class Receipt(BaseModel):
    merchant: str | None = None
    total: float | None = None
    line_items: list[LineItem] = Field(default_factory=list)


class ReceiptOCR(Pipeline):
    steps = [
        ImageIngestor(),
        VisionExtractor(
            schema=Receipt,
            prompt="Extract merchant, total, and line_items from this receipt.",
        ),
    ]


result = ReceiptOCR(runtime=LLAMACPP, model=QWEN3_VL_4B).run("receipt.jpeg")
print(result.model_dump_json(indent=2))

Included

  • Local runtimes: llamacpp, transformers
  • Tasks: summarize, extract, classify, embed, extract_invoice, redact_pii, analyse_sales_call, analyse_video_ad
  • Workflows: InvoiceProcessor, PIIRedactor, SalesCallAnalyser, VideoAdIntelligence
  • Outputs: InvoiceOutput, SalesCallReport, VideoAdReport, RedactedText, Classification

Tool and agent integrations can be added later without changing the core task and workflow layer.

CLI

# Install the runtime or backend extra first
pip install 'aibackends[llamacpp]'
pip install 'aibackends[pii]'

aibackends task extract-invoice --input invoice.pdf --runtime llamacpp --model gemma4-e2b
aibackends task classify --input doc.txt --labels invoice,contract,receipt --runtime llamacpp --model gemma4-e2b
aibackends task redact-pii --input transcript.txt --backend gliner --labels email,phone_number
aibackends pull gemma4-e2b --runtime llamacpp
aibackends check llamacpp --model gemma4-e2b

Full command reference: docs/cli.md.

Docs and Examples

  • docs/usage.md for install, local runtimes, tasks, and workflows
  • docs/concepts.md for task, runtime, backend, model, and workflow terms
  • docs/extending.md for custom runtimes, backends, tasks, and workflows
  • docs/api-reference/index.md for the public API
  • examples/README.md for runnable examples, including local image OCR

Development

python3 -m pip install -e ".[dev]"
python3 -m pytest tests
python3 -m mypy src tests
ruff check .

See CONTRIBUTING.md for contribution guidelines.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

aibackends-0.3.0.tar.gz (96.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

aibackends-0.3.0-py3-none-any.whl (69.0 kB view details)

Uploaded Python 3

File details

Details for the file aibackends-0.3.0.tar.gz.

File metadata

  • Download URL: aibackends-0.3.0.tar.gz
  • Upload date:
  • Size: 96.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: uv/0.9.10 {"installer":{"name":"uv","version":"0.9.10"},"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":null}

File hashes

Hashes for aibackends-0.3.0.tar.gz
Algorithm Hash digest
SHA256 d22a94c6eb0c51e0e5eaa7c6629202d1b231bb3a42ada7b160c36bc85ef3ea74
MD5 c0b634a0f6d4dda45eebc31537882f66
BLAKE2b-256 3942a72a24adf4dee5f7835c2928a97a461f8fb8eada048a8d605fe10dbece0c

See more details on using hashes here.

File details

Details for the file aibackends-0.3.0-py3-none-any.whl.

File metadata

  • Download URL: aibackends-0.3.0-py3-none-any.whl
  • Upload date:
  • Size: 69.0 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: uv/0.9.10 {"installer":{"name":"uv","version":"0.9.10"},"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":null}

File hashes

Hashes for aibackends-0.3.0-py3-none-any.whl
Algorithm Hash digest
SHA256 06ffae813b0b329bf7e163d0560ae4edb2ae8e288e0aa4e28e491187e1517800
MD5 e30e67b3c6ef9b938fb4ce3fe3ce7909
BLAKE2b-256 1e7efa43c8f9974b10fb7a50b0a32395145bb6e95b5cfc5f06e0be13b8774a6e

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page