chatroutes-autobranch 1.2.0

Intelligent branch exploration for LLM-powered applications with branch detection

chatroutes-autobranch

Controlled branching generation for LLM applications

Modern LLM applications often need to explore multiple reasoning paths (tree-of-thought, beam search, multi-agent systems) while staying usable and affordable. chatroutes-autobranch provides clean, standalone primitives for:

• 🔍 Branch Detection – Identify decision points in text (enumerations, disjunctions, conditionals)
• 🎯 Beam Search – Pick the best K candidates by configurable scoring
• 🌈 Diversity Control – Ensure variety via novelty pruning (cosine similarity, MMR)
• 🛑 Smart Stopping – Know when to stop via entropy/information-gain metrics
• 💰 Budget Management – Keep costs predictable with token/time/node caps
• 🔌 Pluggable Design – Swap any component (scorer, embeddings, stopping criteria)

Key Features:

• ✅ Deterministic & reproducible (fixed tie-breaking, seeded clustering)
• ✅ Embedding-agnostic (OpenAI, HuggingFace, or custom)
• ✅ Production-ready (thread-safe, observable, checkpoint/resume)
• ✅ Framework-friendly (works with LangChain, LlamaIndex, or raw LLM APIs)
• ✅ Zero vendor lock-in (MIT License, no cloud dependencies)

---

🚀 Interactive Demos (Try it Now!)

Getting Started Demo (Recommended)

Perfect for first-time users! Learn the fundamentals in 5 minutes:

• ✅ Installation and setup
• ✅ Basic beam search examples
• ✅ Multi-strategy scoring
• ✅ Novelty filtering
• ✅ Complete pipeline with budget control

No setup required - runs entirely in your browser!

Branch Detection Demo (NEW! 🎉)

Analyze text for decision points! Interactive branch detection:

• ✅ Extract branch points from LLM responses
• ✅ Count possible conversation paths
• ✅ Pattern-based detection (no LLM needed)
• ✅ Optional LLM assist for complex cases
• ✅ Try your own text interactively

Creative Writing Scenario (Advanced)

See it in action with a real LLM! Complete creative writing assistant:

• ✅ Full Ollama integration (free, local inference)
• ✅ Multi-turn branching (tree exploration)
• ✅ GPU/CPU performance comparison
• ✅ 4 complete story scenarios

📚 View all notebooks →

---

Quick Start

Install:

pip install chatroutes-autobranch

Basic Usage:

from chatroutes_autobranch import BranchSelector, Candidate
from chatroutes_autobranch.config import load_config

# Load config (or use dict/env vars)
selector = BranchSelector.from_config(load_config("config.yaml"))

# Define parent and candidate branches
parent = Candidate(id="root", text="Explain photosynthesis simply")
candidates = [
    Candidate(id="c1", text="Start with sunlight absorption"),
    Candidate(id="c2", text="Begin with glucose production"),
    Candidate(id="c3", text="Explain chlorophyll's role"),
]

# Select best branches (applies beam → novelty → entropy pipeline)
result = selector.step(parent, candidates)

print(f"Kept: {[c.id for c in result.kept]}")
print(f"Entropy: {result.metrics['entropy']['value']:.2f}")
print(f"Should continue: {result.metrics['entropy']['continue']}")

Config (config.yaml):

beam:
  k: 3  # Keep top 3 by score
  weights: {confidence: 0.4, relevance: 0.3, novelty_parent: 0.2}

novelty:
  method: cosine  # or 'mmr' for Maximal Marginal Relevance
  threshold: 0.85

entropy:
  min_entropy: 0.6  # Stop if diversity drops below 60%

embeddings:
  provider: openai
  model: text-embedding-3-large

---

🔍 Branch Detection (NEW!)

Analyze text to identify decision points before generating branches:

from chatroutes_autobranch import BranchExtractor

# Analyze LLM response for branch points
text = """
Backend options:
1. Flask - lightweight
2. FastAPI - modern
3. Django - full-featured

Database: Postgres or MySQL
"""

extractor = BranchExtractor()
branch_points = extractor.extract(text)

print(f"Found {len(branch_points)} decision points")
# Output: Found 2 decision points

print(f"Max paths: {extractor.count_max_leaves(branch_points)}")
# Output: Max paths: 6 (3 backends × 2 databases)

Features:

• ✅ Deterministic pattern matching - No LLM needed (fast, free)
• ✅ Detects multiple patterns - Enumerations, disjunctions, conditionals
• ✅ Combinatorial counting - Calculate max possible paths (Π ki)
• ✅ Optional LLM assist - Fallback for complex/implicit cases
• ✅ Statistics & analysis - Breakdown by type, complexity metrics

Use Cases:

• Pre-analyze LLM responses before branching
• Count conversation path complexity
• Estimate branching potential from text
• Extract structured choices from unstructured responses

Try it in Colab →

---

Why Use This?

Problem: Exploring multiple LLM reasoning paths (e.g., tree-of-thought) quickly becomes:

• Expensive – Exponential growth of branches drains API budgets
• Redundant – Models generate similar outputs (mode collapse)
• Uncontrolled – No clear stopping criteria (when is "enough" exploration?)

Solution: chatroutes-autobranch gives you:

1. Beam Search to keep only the top-K candidates (quality filtering)
2. Novelty Pruning to remove similar outputs (diversity enforcement)
3. Entropy Stopping to detect when you've explored enough (convergence detection)
4. Budget Limits to cap costs before runaway spending

Result: Controlled, efficient tree exploration with predictable costs.

---

Use Cases

Scenario	Configuration	Benefit
Branch Analysis	BranchExtractor only	Analyze text for decision points, count paths (no generation)
Tree-of-Thought Reasoning	K=5, cosine novelty, entropy stopping	Explore diverse reasoning paths without explosion
Multi-Agent Debate	K=3, MMR novelty (λ=0.3)	Select diverse agent perspectives, avoid redundancy
Code Generation	K=4, high relevance weight	Generate varied solutions, prune duplicates
Creative Writing	K=8, low novelty threshold	High diversity, explore creative space
Factual Q&A	K=2, strict budget	Focus on accuracy, minimal branching

---

Architecture

Two-Phase Workflow:

Phase 1: Branch Detection (Optional, Pre-Analysis)
──────────────────────────────────────────────────
Text → BranchExtractor → Branch Points → Count Max Paths
                       → Statistics
                       → Decision: Generate or Skip?

Phase 2: Branch Selection (Core Pipeline)
──────────────────────────────────────────
Raw Candidates (N)
    ↓
1. Scoring (composite: confidence + relevance + novelty + intent + reward)
    ↓
2. Beam Selection (top K by score, deterministic tie-breaking)
    ↓
3. Novelty Filtering (prune similar via cosine/MMR)
    ↓
4. Entropy Check (compute diversity, decide if should continue)
    ↓
5. Result (kept + pruned + metrics)

Pluggable Components:

• BranchExtractor: Deterministic pattern matching (optional)
• LLMBranchParser: LLM-based extraction (optional fallback)
• Scorer: Composite (built-in) or custom
• EmbeddingProvider: OpenAI, HuggingFace, or custom
• NoveltyFilter: Cosine threshold or MMR
• EntropyStopper: Shannon entropy or custom
• BudgetManager: Token/time/node caps

All components use Protocol (duck typing) – swap any part without touching others.

---

Installation

Minimal:

pip install chatroutes-autobranch

With extras:

# FastAPI service (for TypeScript/other languages)
pip install chatroutes-autobranch[service]

# HuggingFace local embeddings
pip install chatroutes-autobranch[hf]

# FAISS for large-scale similarity (1000+ candidates)
pip install chatroutes-autobranch[faiss]

# All features
pip install chatroutes-autobranch[all]

---

Documentation

📘 Full Specification – Complete API reference, algorithms, examples, and troubleshooting

Key Sections:

• Philosophy & Design – Core principles
• Pluggable Interfaces – Protocols & implementations
• Configuration – YAML/JSON/env setup
• Examples – Single-step & multi-generation
• Tuning Guide – How to choose K
• Common Failures – Troubleshooting

---

Examples

Multi-Generation Tree Exploration

from collections import deque
import time

# User provides LLM generation function
def my_llm_generate(parent: Candidate, n: int) -> list[Candidate]:
    # Your LLM call here (OpenAI, Anthropic, etc.)
    responses = llm_api.generate(parent.text, n=n)
    return [Candidate(id=f"{parent.id}_{i}", text=r) for i, r in enumerate(responses)]

# Setup
selector = BranchSelector.from_config(load_config("config.yaml"))
budget_manager = BudgetManager(Budget(max_nodes=50, max_tokens=20000))

# Tree exploration
queue = deque([root_candidate])
while queue:
    current = queue.popleft()
    children = my_llm_generate(current, n=5)

    # Check budget before selection
    if not budget_manager.admit(n_new=5, est_tokens=1000, est_ms=2000):
        break

    # Select best branches
    result = selector.step(current, children)
    budget_manager.update(actual_tokens=1200, actual_ms=1800)

    # Continue with kept candidates
    queue.extend(result.kept)

    # Stop if entropy is low (converged)
    if not result.metrics["entropy"]["continue"]:
        break

Custom Scorer

from chatroutes_autobranch import Scorer, Candidate, ScoredCandidate

class DomainScorer(Scorer):
    def score(self, parent: Candidate, candidates: list[Candidate]) -> list[ScoredCandidate]:
        scored = []
        for c in candidates:
            # Custom logic: prefer longer, detailed responses
            detail_score = min(len(c.text) / 1000, 1.0)
            scored.append(ScoredCandidate(id=c.id, text=c.text, score=detail_score))
        return scored

# Use in pipeline
beam = BeamSelector(k=3, scorer=DomainScorer())
selector = BranchSelector(beam, novelty, entropy, budget)

FastAPI Service (for TypeScript/other languages)

# server.py
from fastapi import FastAPI
from chatroutes_autobranch import BranchSelector
from chatroutes_autobranch.config import load_config_from_file

app = FastAPI()
_config = load_config_from_file("config.yaml")

@app.post("/select")
async def select(parent: dict, candidates: list[dict]):
    # Create fresh selector per request (thread-safe)
    selector = BranchSelector.from_config(_config)
    result = selector.step(
        Candidate(**parent),
        [Candidate(**c) for c in candidates]
    )
    return {
        "kept": [{"id": c.id, "score": c.score} for c in result.kept],
        "metrics": result.metrics
    }

# Run: uvicorn server:app

TypeScript client:

const response = await fetch('http://localhost:8000/select', {
  method: 'POST',
  body: JSON.stringify({ parent, candidates })
});
const { kept, metrics } = await response.json();

---

Features

Beam Search

• Top-K selection by composite scoring
• Deterministic tie-breaking (lexicographic ID ordering)
• Configurable weights: confidence, relevance, novelty, intent alignment, historical reward

Novelty Pruning

• Cosine similarity: Remove candidates above threshold (e.g., 0.85)
• MMR (Maximal Marginal Relevance): Balance relevance vs diversity with λ parameter
• Preserves score ordering (best candidates kept first)

Entropy-Based Stopping

• Shannon entropy on K-means clusters of embeddings
• Delta-entropy tracking (stop if change < epsilon)
• Handles edge cases (0, 1, 2 candidates)
• Normalized to [0,1] scale

Budget Management

• Caps: max_nodes, max_tokens, max_ms
• Modes: strict (raise on exceeded) or soft (return False, allow fallback)
• Pre-admit: Check budget before generation
• Post-update: Record actual usage for rolling averages

Observability

• Structured JSON logging (PII-safe by default)
• OpenTelemetry spans (optional)
• Rich metrics per step (kept/pruned counts, scores, entropy, budget usage)

Checkpointing

• Serialize selector state (entropy history, budget snapshot)
• Resume from checkpoint (pause/resume tree exploration)
• Schema versioning for backward compatibility

---

Integrations

LangChain:

from langchain.chains import LLMChain
from chatroutes_autobranch import Candidate, BranchSelector

def generate_and_select(query: str, chain: LLMChain, selector: BranchSelector):
    # Generate N candidates via LangChain
    responses = chain.generate([{"query": query}] * 5)
    candidates = [Candidate(id=f"c{i}", text=r.text) for i, r in enumerate(responses.generations[0])]

    # Select best
    parent = Candidate(id="root", text=query)
    result = selector.step(parent, candidates)
    return result.kept

LlamaIndex: Similar pattern using QueryEngine.query() for generation

Raw APIs (OpenAI, Anthropic): See multi-generation example

---

Performance

Benchmarks (M1 Max, OpenAI embeddings):

Candidates	Beam K	Latency (p50)	Bottleneck
10	3	240ms	Embedding API
50	5	520ms	Embedding API
100	10	1.1s	Novelty O(N²)
500	10	4.2s	Use FAISS

Optimization tips:

• Use local embeddings (HuggingFace) for <100ms latency
• Enable FAISS for 100+ candidates
• Batch embedding calls (batch_size: 64 in config)
• Global embedding cache for repeated candidates

---

Development

Setup:

git clone https://github.com/chatroutes/chatroutes-autobranch
cd chatroutes-autobranch
pip install -e .[dev]

Run tests:

pytest tests/
pytest tests/ -v --cov=chatroutes_autobranch  # With coverage

Type checking:

mypy src/

Formatting:

black src/ tests/
ruff check src/ tests/

Benchmarks:

pytest bench/ --benchmark-only

---

Contributing

We welcome contributions! Please see our contributing guidelines.

Areas we'd love help with:

• Additional novelty algorithms (DPP, k-DPP)
• More embedding providers (Cohere, Voyage AI)
• Adaptive K scheduling (auto-tune beam width)
• Tree visualization tools
• More examples (specific domains)

How to contribute:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/amazing-feature)
3. Make your changes with tests
4. Run tests and type checking
5. Submit a Pull Request

---

Roadmap

• v1.0.0 ✅ RELEASED (January 2025): Core components, beam search, MMR novelty, cosine filtering, entropy stopping, budget management, full test suite
• v1.1.0 (Q2 2025): FAISS support for large-scale similarity, adaptive K scheduling
• v1.2.0 (Q3 2025): Tree visualization tools, FastAPI service for multi-language support
• v1.3.0 (Q4 2025): Async/await support, cluster-aware pruning
• v2.0.0 (Q1 2026): gRPC service, TypeScript SDK, breaking API improvements

---

FAQ

Q: Do I need ChatRoutes cloud to use this? A: No. This library is standalone and has zero cloud dependencies. Use it with any LLM provider.

Q: Can I use this with TypeScript/JavaScript? A: Yes. Run the FastAPI service and call via HTTP. Native TS SDK planned for v2.0.0.

Q: How do I choose beam width K? A: Start with K=3-5. Use budget formula: K ≈ (budget/tokens_per_branch)^(1/depth). See tuning guide.

Q: What if all candidates get pruned by novelty? A: Lower threshold (e.g., 0.75) or switch to MMR. See troubleshooting.

Q: Is this deterministic? A: Yes, with fixed random seeds and deterministic tie-breaking. See tests.

---

License

MIT License - see LICENSE file for details.

---

Acknowledgements

Inspired by research in beam search, diverse selection (MMR, DPP), and LLM orchestration patterns. Built to be practical, swappable, and friendly for contributors.

Special thanks to the open-source community for tools and inspiration: LangChain, LlamaIndex, HuggingFace Transformers, FAISS, and the broader LLM ecosystem.

---

Links

• Documentation: Full Specification
• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Changelog: CHANGELOG.md
• PyPI: pypi.org/project/chatroutes-autobranch

---

Built with ❤️ by the ChatRoutes team. Open to the community.