silkloom-core 1.0.0

SilkLoom Core: minimal stateful batch engine for LLM and VLM workloads

SilkLoom Core

中文 | English

SilkLoom Core V1.0.0 is a minimal, stateful LLM batch engine.

The public surface is intentionally small:

• LLMTask
• ResultSet
• TaskResult

This README is split into two parts:

1. User Guide: installation, input formats, prompt rules, and examples.
2. API Reference: constructor arguments, method signatures, and returned objects.

Prompt templates use strict Jinja2 syntax. user_prompt and system_prompt render against each input item, so template variables must match the keys in that item's dictionary. Missing variables raise an error instead of rendering as empty text. For a plain string list, SilkLoom wraps each item as {"text": "..."}.

Install

pip install silkloom-core

From source:

git clone https://github.com/LeLiu-GeoAI/silkloom-core.git
cd silkloom-core
pip install -e .

User Guide

Quick Start

from openai import OpenAI
from silkloom_core import LLMTask

client = OpenAI(api_key="your_key")

task = LLMTask(
    model="gpt-4o-mini",
    user_prompt="Translate into English: {{ text }}",
    client=client,
)

results = task.map(["你好", "今天天气不错"])
print(results[0])
print(results.success_count, results.failed_count)

Input Formats

LLMTask.map() accepts three common input shapes:

• list[str]: each string is wrapped as {"text": ...}
• list[dict]: each dict becomes one prompt context
• pandas.DataFrame: each row becomes one prompt context and the column names become template variables

Dictionary list example:

from silkloom_core import LLMTask

task = LLMTask(
    model="gpt-4o-mini",
    user_prompt="Extract name and intent from text: {{ text }}",
)

results = task.map([
    {"text": "My name is Alice. I want a refund."},
    {"text": "Bob asks about delivery."},
])

Pandas DataFrame

Each DataFrame row is treated as one input item, and the column names are available as template variables.

import pandas as pd
from silkloom_core import LLMTask

df = pd.DataFrame(
    [
        {"text": "Urban heat island is intensifying.", "lang": "en"},
        {"text": "城市更新需要兼顾公平。", "lang": "zh"},
    ]
)

task = LLMTask(
    model="gpt-4o-mini",
    user_prompt="Rewrite the following {{ lang }} text: {{ text }}",
)

results = task.map(df)

Prompt Template Rules

Template variables must match the keys in the input context.

task = LLMTask(
    model="gpt-4o-mini",
    user_prompt="Rewrite the following {{ lang }} text: {{ text }}",
)

For a DataFrame, this row exposes text and lang to the template:

{"text": "Urban heat is rising.", "lang": "en"}

Structured Output

from pydantic import BaseModel
from silkloom_core import LLMTask


class ExtractInfo(BaseModel):
    name: str
    intent: str


task = LLMTask(
    model="gpt-4o-mini",
    user_prompt="Extract name and intent from text: {{ text }}",
    response_model=ExtractInfo,
)

results = task.map([
    {"text": "My name is Alice. I want a refund."},
    {"text": "Bob asks about delivery."},
])

print(results[0].name)

GLM and Ollama

GLM-4-Flash

import os
from openai import OpenAI
from silkloom_core import LLMTask

glm_client = OpenAI(
    api_key=os.environ["ZHIPUAI_API_KEY"],
    base_url="https://open.bigmodel.cn/api/paas/v4/",
)

task = LLMTask(
    model="glm-4-flash",
    user_prompt="Summarize this text: {{ text }}",
    client=glm_client,
)

results = task.map(["Urban renewal should balance efficiency and equity."])

Ollama

from openai import OpenAI
from silkloom_core import LLMTask

ollama_client = OpenAI(
    api_key="ollama",
    base_url="http://localhost:11434/v1",
)

task = LLMTask(
    model="qwen2.5:7b",
    user_prompt="Rewrite in academic tone: {{ text }}",
    client=ollama_client,
)

results = task.map(["Traffic is usually worst in evening peak."])

Multimodal Input

Pass image sources in images (supports local path, URL, or base64/data URI):

from silkloom_core import LLMTask

task = LLMTask(
    model="gpt-4o",
    user_prompt="Describe these images and answer: {{ text }}",
)

results = task.map([
    {
        "text": "What is shown?",
        "images": ["./pic1.jpg", "https://example.com/pic2.png"],
    }
])

Resumability

map supports resumability with SQLite via db_path + run_id:

results = task.map(
    [{"text": "a"}, {"text": "b"}],
    db_path="my_run.db",
    run_id="demo_001",
    workers=5,
)

Running again with the same run_id reuses successful records.

Exporting Results

ResultSet supports in-memory access and file export:

results.run_id
results.success_count
results.failed_count
results.total_tokens
results.errors
results[0]
results.export_jsonl("out.jsonl")
results.export_csv("out.csv", flatten=True)

API Reference

LLMTask

Constructor:

LLMTask(
    model: str,
    user_prompt: str,
    system_prompt: str | None = None,
    response_model: type[BaseModel] | None = None,
    max_retries: int = 3,
    client: Any | None = None,
)

Arguments:

• model: target model name, such as gpt-4o-mini
• user_prompt: required Jinja2 template for the user message
• system_prompt: optional Jinja2 template for the system message
• response_model: optional Pydantic model for structured output parsing
• max_retries: number of attempts for one item
• client: optional OpenAI-compatible client; defaults to the official client

Method:

map(sequence, db_path=".silkloom_cache.db", run_id=None, workers=5) -> ResultSet

Accepted inputs:

• list[str]
• list[dict]
• pandas.DataFrame

ResultSet

ResultSet behaves like a sequence aligned with the input order.

Properties:

• run_id
• success_count
• failed_count
• total_tokens
• errors

Methods:

• results[0]: returns the result at the same index as the input
• export_jsonl(path): write successful results to JSONL
• export_csv(path, flatten=False, include_usage=True): write a CSV export

TaskResult

Each raw task result contains:

• is_success
• data
• error
• usage
• input_data

License

MIT