RuleChef

Learn rule-based models from examples using LLM-powered synthesis.
Replace expensive LLM calls with fast, deterministic, inspectable rules.

---

What is RuleChef?

RuleChef learns regex, Python code, and spaCy patterns from labeled examples using LLM-powered synthesis. You provide examples, RuleChef generates rules, and those rules run locally without any LLM at inference time.

Why rules instead of LLMs?

• Cost: Rules cost nothing to run. No API calls, no tokens.
• Latency: Sub-millisecond per query vs hundreds of ms for LLM calls.
• Determinism: Same input always produces the same output.
• Inspectability: You can read, edit, and debug every rule.
• No drift: Rules don't change unless you change them.

Results

On the Text Anonymization Benchmark (real court decisions, PII extraction), rules learned by RuleChef beat both prompting the same LLM and a dedicated neural extractor on format-structured entity types — at a fraction of the cost:

	Format-type F1	ms/doc
RuleChef (rules only)	78.7	0.6
LLM prompting (same model)	74.8	~1500
GLiNER2 (schema)	74.4	190

The LLM stays ahead on purely semantic types — rules are the high-precision tier, not a full replacement. In observation mode, rules learned from watching just 50 LLM calls replace 48% of subsequent calls at 96% precision. Details and per-type numbers: benchmarks/results/.

Installation

pip install rulechef

Extras:

pip install rulechef[grex]     # Regex pattern suggestions from examples
pip install rulechef[spacy]    # spaCy token/dependency matcher patterns
pip install rulechef[agentic]  # LLM-powered coordinator for adaptive learning
pip install rulechef[all]      # Everything

Quick Start

Extraction

Extract answer spans from text:

from openai import OpenAI
from rulechef import RuleChef, Task, TaskType

client = OpenAI()
task = Task(
    name="Q&A Extraction",
    description="Extract answer spans from context",
    input_schema={"question": "str", "context": "str"},
    output_schema={"spans": "List[Span]"},
    type=TaskType.EXTRACTION,
)

chef = RuleChef(task, client)

chef.add_example(
    {"question": "When?", "context": "Built in 1991"},
    {"spans": [{"text": "1991", "start": 9, "end": 13}]}
)
chef.add_example(
    {"question": "When?", "context": "Released in 2005"},
    {"spans": [{"text": "2005", "start": 12, "end": 16}]}
)

chef.learn_rules()

result = chef.extract({"question": "When?", "context": "Founded in 1997"})
print(result)  # {"spans": [{"text": "1997", ...}]}

Named Entity Recognition (NER)

from pydantic import BaseModel
from typing import List, Literal

class Entity(BaseModel):
    text: str
    start: int
    end: int
    type: Literal["DRUG", "DOSAGE", "CONDITION"]

class NEROutput(BaseModel):
    entities: List[Entity]

task = Task(
    name="Medical NER",
    description="Extract drugs, dosages, and conditions",
    input_schema={"text": "str"},
    output_schema=NEROutput,
    type=TaskType.NER,
)

chef = RuleChef(task, client)
chef.add_example(
    {"text": "Take Aspirin 500mg for headache"},
    {"entities": [
        {"text": "Aspirin", "start": 5, "end": 12, "type": "DRUG"},
        {"text": "500mg", "start": 13, "end": 18, "type": "DOSAGE"},
        {"text": "headache", "start": 23, "end": 31, "type": "CONDITION"},
    ]}
)
chef.learn_rules()

Classification

task = Task(
    name="Intent Classification",
    description="Classify banking customer queries",
    input_schema={"text": "str"},
    output_schema={"label": "str"},
    type=TaskType.CLASSIFICATION,
    text_field="text",
)

chef = RuleChef(task, client)
chef.add_example({"text": "what is the exchange rate?"}, {"label": "exchange_rate"})
chef.add_example({"text": "I want to know the rates"}, {"label": "exchange_rate"})
chef.add_example({"text": "my card hasn't arrived"}, {"label": "card_arrival"})

chef.learn_rules()
result = chef.extract({"text": "current exchange rate please"})
print(result)  # {"label": "exchange_rate"}

Transformation

task = Task(
    name="Invoice Parser",
    description="Extract company and amount from invoices",
    input_schema={"text": "str"},
    output_schema={"company": "str", "amount": "str"},
    type=TaskType.TRANSFORMATION,
)

chef = RuleChef(task, client)
chef.add_example(
    {"text": "Invoice from Acme Corp for $1,500.00"},
    {"company": "Acme Corp", "amount": "$1,500.00"}
)
chef.learn_rules()

Core Concepts

Task Types

Type	Output	Use Case
EXTRACTION	{"spans": [...]}	Find text spans (untyped)
NER	{"entities": [...]}	Find typed entities with labels
CLASSIFICATION	{"label": "..."}	Classify text into categories
TRANSFORMATION	Custom dict	Extract structured fields

Rule Formats

Format	Best For	Example
RuleFormat.REGEX	Keyword patterns, structured text	\b\d{4}\b
RuleFormat.CODE	Complex logic, multi-field extraction	def extract(input_data): ...
RuleFormat.SPACY	Linguistic patterns, POS/dependency	[{"POS": "PROPN", "OP": "+"}]

from rulechef import RuleFormat

# Only generate regex rules (fastest, most portable)
chef = RuleChef(task, client, allowed_formats=[RuleFormat.REGEX])

# Only code rules (most flexible)
chef = RuleChef(task, client, allowed_formats=[RuleFormat.CODE])

Buffer-First Architecture

Examples go to a buffer first, then get committed to the dataset during learn_rules(). This enables batch learning and coordinator-driven decisions:

chef.add_example(input1, output1)   # Goes to buffer
chef.add_example(input2, output2)   # Goes to buffer
chef.add_correction(input3, wrong_output, correct_output)  # High-priority signal

chef.learn_rules()  # Buffer -> Dataset -> Synthesis -> Refinement

Corrections & Feedback

Corrections are the highest-value training signal -- they show exactly where the current rules fail:

result = chef.extract({"text": "some input"})
# Result was wrong! Correct it:
chef.add_correction(
    {"text": "some input"},
    model_output=result,
    expected_output={"label": "correct_label"},
    feedback="The rule matched too broadly"
)

# Task-level guidance
chef.add_feedback("Drug names always follow 'take' or 'prescribe'")

# Rule-level feedback
chef.add_feedback("This rule is too broad", level="rule", target_id="rule_id")

chef.learn_rules()  # Re-learns with corrections prioritized

Evaluation

RuleChef includes built-in evaluation with entity-level precision, recall, and F1:

# Dataset-level evaluation
eval_result = chef.evaluate()
# Prints: Exact match, micro/macro P/R/F1, per-class breakdown

# Per-rule evaluation (find dead or harmful rules)
metrics = chef.get_rule_metrics()
# Shows: per-rule TP/FP/FN, sample matches, identifies dead rules

# Delete a bad rule
chef.delete_rule("rule_id")

Inspect Your Rules

The rules are the model — so you can read them. Generate a browsable report of any ruleset against labeled data: per-rule precision, and every true/false positive highlighted in context.

rulechef-report --rules my_rules.json --data gold.jsonl --out report.html

Load a saved ruleset back any time — from a dataset file, a benchmark result, or a bare list of rule dicts:

chef.load_rules("benchmarks/results/results_extract_tab.ckpt_rulechef.json")
chef.extract({"text": "filed under no. 36244/06 in 2006"})   # runs rules, no LLM

See benchmarks/INSPECTING_RULES.md.

Advanced Features

Synthesis Strategy

For multi-class tasks, RuleChef can synthesize rules one class at a time for better coverage:

# Auto-detect (default): per-class if >1 class, bulk otherwise
chef = RuleChef(task, client, synthesis_strategy="auto")

# Force per-class synthesis
chef = RuleChef(task, client, synthesis_strategy="per_class")

# Force single-prompt bulk synthesis
chef = RuleChef(task, client, synthesis_strategy="bulk")

Agentic Coordinator

The AgenticCoordinator uses LLM calls to guide the refinement loop, focusing on weak classes:

from rulechef import RuleChef, AgenticCoordinator

coordinator = AgenticCoordinator(client, model="gpt-4o-mini")
chef = RuleChef(task, client, coordinator=coordinator)

chef.learn_rules(max_refinement_iterations=10)
# Coordinator analyzes per-class metrics each iteration,
# tells the synthesis prompt which classes to focus on,
# and stops early when performance plateaus.

Rule Pruning

With prune_after_learn=True, the agentic coordinator audits rules after learning -- merging redundant rules and removing pure noise. A safety net reverts if F1 drops:

coordinator = AgenticCoordinator(client, prune_after_learn=True)
chef = RuleChef(task, client, coordinator=coordinator)

chef.learn_rules()
# After synthesis+refinement:
# 1. LLM analyzes rules + per-rule metrics
# 2. Merges similar patterns (e.g. two regexes → one)
# 3. Removes precision=0 rules (pure false positives)
# 4. Re-evaluates — reverts if F1 drops

In the CLI: learn --agentic --prune.

Incremental Patching

After the initial learn, you can patch existing rules without full re-synthesis:

chef.learn_rules()           # Initial synthesis
chef.add_correction(...)     # Add corrections
chef.learn_rules(incremental_only=True)  # Patch, don't re-synthesize

Observation Mode

Collect training data from any LLM -- no task definition needed:

# Works with any LLM provider (Anthropic, Groq, local models, etc.)
chef = RuleChef(client=client, model="gpt-4o-mini")  # No task needed
chef.add_observation({"text": "what's the exchange rate?"}, {"label": "exchange_rate"})
chef.learn_rules()  # Auto-discovers the task schema

For raw LLM interactions where you don't know the schema:

chef.add_raw_observation(
    messages=[{"role": "user", "content": "classify: what's the rate?"}],
    response="exchange_rate",
)
chef.learn_rules()  # Discovers task + maps observations + learns rules

For OpenAI-compatible clients, auto-capture with monkey-patching:

wrapped = chef.start_observing(openai_client, auto_learn=False)
response = wrapped.chat.completions.create(...)  # Observed automatically
chef.learn_rules()
chef.stop_observing()

Pydantic Output Schemas

Use Pydantic models for type-safe, validated outputs with automatic label extraction:

from pydantic import BaseModel
from typing import List, Literal

class Entity(BaseModel):
    text: str
    start: int
    end: int
    type: Literal["PERSON", "ORG", "LOCATION"]

class Output(BaseModel):
    entities: List[Entity]

task = Task(..., output_schema=Output, type=TaskType.NER)
# RuleChef automatically discovers labels: ["PERSON", "ORG", "LOCATION"]

grex: Regex Pattern Suggestions

When use_grex=True (default), grex analyzes your training examples and adds regex pattern hints to the synthesis prompt. The LLM sees concrete patterns alongside the raw examples, producing better rules — especially for structured data like dates, IDs, and amounts:

DATA EVIDENCE FROM TRAINING:
- DATE (5 unique): "2024-01-15", "2024-02-28", "2023-12-01", ...
  Exact pattern: (2023\-12\-01|2024\-01\-15|2024\-02\-28|...)
  Structural pattern: \d{4}\-\d{2}\-\d{2}

Install with pip install rulechef[grex]. Disable with use_grex=False.

Benchmarks

All numbers, harnesses, and learned rulesets are committed under benchmarks/ — every reported result traces to a script and a results JSON. Highlights:

• TAB anonymization (real court decisions): format-type F1 78.7 vs 74.8 (LLM prompting) vs 74.4 (GLiNER2); on the official test split the rules recover 83% of direct identifiers at 5 ms per document on CPU.
• Ablation: one-shot rule prompting gets 7.4 F1 on semantic types; the refinement loop with holdout acceptance reaches 44.4 — the pipeline, not the prompt, does the work.
• Feedback repair: three sentences of plain-English rule feedback lift a broken rule family from 5.7 to 35.6 F1 in one incremental round.
• Banking77 (5-class, 5-shot): 97.6% precision at 61% recall (75.1 micro-F1, 0.17 ms/query) — rules answer or abstain, never guess.

CLI

Interactive CLI for quick experimentation across all task types:

export OPENAI_API_KEY=your_key
rulechef

The CLI walks you through a setup wizard (task name, type, labels, model, base URL) and drops you into a command loop:

Commands:
  add        Add a training example
  correct    Add a correction
  extract    Run extraction on input
  learn      Learn rules (--iterations N, --incremental, --agentic, --prune)
  evaluate   Evaluate rules against dataset
  rules      List learned rules (rules <id> for detail)
  delete     Delete a rule by ID
  feedback   Add feedback (task/rule level)
  generate   Generate synthetic examples with LLM
  stats      Show dataset statistics
  help       Show commands
  quit       Exit

Works with any OpenAI-compatible API (Groq, Together, Ollama, etc.) via the base URL prompt.

Web App

RuleChef includes a web UI (FastAPI + React) for interactive rule learning. Upload data, learn rules, see highlighted entities, and correct mistakes — all in the browser.

pip install -e ".[app]"
uvicorn api.main:app --reload --port 8000

cd frontend && npm install && npm run dev

See the Web App guide for full setup and usage.

License

Apache 2.0 -- see LICENSE.