semantic-frame 0.2.1

Semantic compression layer for numerical data - token-efficient LLM context with Anthropic Advanced Tool Use support

Semantic Frame

Token-efficient semantic compression for numerical data.

Semantic Frame converts raw numerical data (NumPy, Pandas, Polars) into natural language descriptions optimized for LLM consumption. Instead of sending thousands of data points to an AI agent, send a 50-word semantic summary.

The Problem

LLMs are terrible at arithmetic. When you send raw data like [100, 102, 99, 101, 500, 100, 98] to GPT-4 or Claude:

• Token waste: 1000 data points = ~2000 tokens
• Hallucination risk: LLMs guess trends instead of calculating them
• Context overflow: Large datasets fill the context window

The Solution

Semantic Frame provides deterministic analysis using NumPy, then translates results into token-efficient narratives:

from semantic_frame import describe_series
import pandas as pd

data = pd.Series([100, 102, 99, 101, 500, 100, 98])
print(describe_series(data, context="Server Latency (ms)"))

Output:

The Server Latency (ms) data shows a flat/stationary pattern with stable
variability. 1 anomaly detected at index 4 (value: 500.00).
Baseline: 100.00 (range: 98.00-500.00).

Result: 95%+ token reduction, zero hallucination risk.

Installation

pip install semantic-frame

Or with uv:

uv add semantic-frame

Quick Start

Analyze a Series

from semantic_frame import describe_series
import numpy as np

# Works with NumPy arrays
data = np.array([10, 20, 30, 40, 50, 60, 70, 80, 90, 100])
result = describe_series(data, context="Daily Sales")
print(result)
# "The Daily Sales data shows a rapidly rising pattern with moderate variability..."

Analyze a DataFrame

from semantic_frame import describe_dataframe
import pandas as pd

df = pd.DataFrame({
    'cpu': [40, 42, 41, 95, 40, 41],
    'memory': [60, 61, 60, 60, 61, 60],
})

results = describe_dataframe(df, context="Server Metrics")
print(results['cpu'].narrative)
# "The Server Metrics - cpu data shows a flat/stationary pattern..."

Get Structured Output

result = describe_series(data, output="full")

print(result.trend)          # TrendState.RISING_SHARP
print(result.volatility)     # VolatilityState.MODERATE
print(result.anomalies)      # [AnomalyInfo(index=4, value=500.0, z_score=4.2)]
print(result.compression_ratio)  # 0.95

JSON Output for APIs

result = describe_series(data, output="json")
# Returns dict ready for JSON serialization

Supported Data Types

• NumPy: np.ndarray
• Pandas: pd.Series, pd.DataFrame
• Polars: pl.Series, pl.DataFrame
• Python: list

Analysis Features

Feature	Method	Output
Trend	Linear regression slope	RISING_SHARP, RISING_STEADY, FLAT, FALLING_STEADY, FALLING_SHARP
Volatility	Coefficient of variation	COMPRESSED, STABLE, MODERATE, EXPANDING, EXTREME
Anomalies	Z-score / IQR adaptive	Index, value, z-score for each outlier
Seasonality	Autocorrelation	NONE, WEAK, MODERATE, STRONG
Distribution	Skewness + Kurtosis	NORMAL, LEFT_SKEWED, RIGHT_SKEWED, BIMODAL, UNIFORM
Step Change	Baseline shift detection	NONE, STEP_UP, STEP_DOWN
Data Quality	Missing value %	PRISTINE, GOOD, SPARSE, FRAGMENTED

LLM Integration

System Prompt Injection

from semantic_frame.interfaces import format_for_system_prompt

result = describe_series(data, output="full")
prompt = format_for_system_prompt(result)
# Returns formatted context block for system prompts

LangChain Tool Output

from semantic_frame.interfaces import format_for_langchain

output = format_for_langchain(result)
# {"output": "narrative...", "metadata": {...}}

Multi-Column Agent Context

from semantic_frame.interfaces import create_agent_context

results = describe_dataframe(df)
context = create_agent_context(results)
# Combined narrative for all columns with attention flags

Framework Integrations

Anthropic Claude (Native Tool Use)

pip install semantic-frame[anthropic]

import anthropic
from semantic_frame.integrations.anthropic import get_anthropic_tool, handle_tool_call

client = anthropic.Anthropic()
tool = get_anthropic_tool()

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    tools=[tool],
    messages=[{"role": "user", "content": "Analyze this sales data: [100, 120, 115, 500, 118]"}]
)

# Handle tool use in response
for block in response.content:
    if block.type == "tool_use" and block.name == "semantic_analysis":
        result = handle_tool_call(block.input)
        print(result)

LangChain

pip install semantic-frame[langchain]

from semantic_frame.integrations.langchain import get_semantic_tool

tool = get_semantic_tool()
# Use as a LangChain BaseTool in your agent

CrewAI

pip install semantic-frame[crewai]

from semantic_frame.integrations.crewai import get_crewai_tool

tool = get_crewai_tool()
# Use with CrewAI agents

MCP (Model Context Protocol)

pip install semantic-frame[mcp]

Run the MCP server:

mcp run semantic_frame.integrations.mcp:mcp

Exposes describe_data tool for MCP clients like:

• ElizaOS: TypeScript-based agent framework
• Claude Desktop: Anthropic's desktop app
• Claude Code: Anthropic's CLI for Claude
• Any MCP-compatible client

Claude Code

Add Semantic Frame as a native tool in Claude Code:

# Install MCP dependencies
pip install semantic-frame[mcp]

# Add MCP server to Claude Code
claude mcp add semantic-frame -- uv run --project /path/to/semantic-frame mcp run /path/to/semantic-frame/semantic_frame/integrations/mcp.py

# Restart Claude Code, then verify connection
claude mcp list
# semantic-frame: ... - ✓ Connected

Once configured, ask Claude to analyze data and it will use the describe_data tool automatically.

Advanced Tool Use (Beta)

Semantic Frame supports Anthropic's Advanced Tool Use features for efficient tool orchestration in complex agent workflows.

Features

Feature	Benefit	API
Input Examples	+18% parameter accuracy	Included by default
Tool Search	1000+ tools without context bloat	defer_loading=True
Programmatic Calling	Batch analysis via code execution	allowed_callers=["code_execution"]

Quick Start (Advanced)

import anthropic
from semantic_frame.integrations.anthropic import get_advanced_tool, handle_tool_call

client = anthropic.Anthropic()
tool = get_advanced_tool()  # All advanced features enabled

response = client.beta.messages.create(
    betas=["advanced-tool-use-2025-11-20"],
    model="claude-sonnet-4-5-20250929",
    max_tokens=4096,
    tools=[
        {"type": "tool_search_tool_regex_20251119", "name": "tool_search"},
        {"type": "code_execution_20250825", "name": "code_execution"},
        tool,
    ],
    messages=[{"role": "user", "content": "Analyze all columns in this dataset..."}]
)

Configuration Options

from semantic_frame.integrations.anthropic import (
    get_anthropic_tool,          # Standard (includes examples)
    get_tool_for_discovery,      # For Tool Search
    get_tool_for_batch_processing,  # For code execution
    get_advanced_tool,           # All features enabled
)

MCP Batch Analysis

from semantic_frame.integrations.mcp import describe_batch

# Analyze multiple series in one call
result = describe_batch(
    datasets='{"cpu": [45, 47, 95, 44], "memory": [60, 61, 60, 61]}',
)

See docs/advanced-tool-use.md for complete documentation.

Use Cases

Crypto Trading

btc_prices = pd.Series(hourly_btc_prices)
insight = describe_series(btc_prices, context="BTC/USD Hourly")
# "The BTC/USD Hourly data shows a rapidly rising pattern with extreme variability.
#  Step up detected at index 142. 2 anomalies detected at indices 89, 203."

DevOps Monitoring

cpu_data = pd.Series(cpu_readings)
insight = describe_series(cpu_data, context="CPU Usage %")
# "The CPU Usage % data shows a flat/stationary pattern with stable variability
#  until index 850, where a critical anomaly was detected..."

Sales Analytics

sales = pd.Series(daily_sales)
insight = describe_series(sales, context="Daily Revenue")
# "The Daily Revenue data shows a steadily rising pattern with weak cyclic pattern
#  detected. Baseline: $12,450 (range: $8,200-$18,900)."

IoT Sensor Data

temps = pl.Series("temperature", sensor_readings)
insight = describe_series(temps, context="Machine Temperature (C)")
# "The Machine Temperature (C) data is expanding with extreme outliers.
#  3 anomalies detected at indices 142, 156, 161."

API Reference

describe_series(data, context=None, output="text")

Analyze a single data series.

Parameters:

• data: Input data (NumPy array, Pandas Series, Polars Series, or list)
• context: Optional label for the data (appears in narrative)
• output: Format - "text" (string), "json" (dict), or "full" (SemanticResult)

Returns: Semantic description in requested format.

describe_dataframe(df, context=None)

Analyze all numeric columns in a DataFrame.

Parameters:

• df: Pandas or Polars DataFrame
• context: Optional prefix for column context labels

Returns: Dict mapping column names to SemanticResult objects.

SemanticResult

Full analysis result with:

• narrative: Human-readable text description
• trend: TrendState enum
• volatility: VolatilityState enum
• data_quality: DataQuality enum
• anomaly_state: AnomalyState enum
• anomalies: List of AnomalyInfo objects
• seasonality: Optional SeasonalityState
• distribution: Optional DistributionShape
• step_change: Optional StructuralChange (STEP_UP, STEP_DOWN, NONE)
• step_change_index: Optional int (index where step change occurred)
• profile: SeriesProfile with statistics
• compression_ratio: Token reduction ratio

Development

# Clone and install
git clone https://github.com/yourusername/semantic-frame
cd semantic-frame
uv sync

# Run tests
uv run pytest

# Run with coverage
uv run pytest --cov=semantic_frame

License

MIT License - see LICENSE file.