Stop tuning RAG pipelines. Let them evolve.

SIARE is the first self-improving RAG engine that treats pipeline configuration as evolvable genetic material. Instead of manually tuning prompts, retrieval strategies, and agent topologies, SIARE uses Quality-Diversity optimization to automatically discover and maintain diverse high-performing multi-agent RAG strategies.

Building RAG systems today means endless iteration:

• Tweak prompts → benchmark → repeat
• Try different chunking strategies → benchmark → repeat
• Adjust retrieval parameters → benchmark → repeat
• Add more agents → debug interactions → repeat

This process is expensive, brittle, and never-ending as your data and requirements change.

SIARE treats your RAG pipeline as a living system that evolves:

1. Define your goals - accuracy, latency, cost, or custom metrics
2. Let evolution work - SIARE mutates prompts, tools, and even agent topology
3. Get diverse solutions - Quality-Diversity ensures you have multiple good options
4. Adapt continuously - As your data changes, your pipeline evolves

pip install siare

# Initialize a new project
siare init

# Run evolution (10 generations)
siare evolve

# Query your evolved pipeline
siare run "How do I reset my password?"

┌─────────────────────────────────────────────────────────────────┐
│                      SIARE Evolution Loop                        │
├─────────────────────────────────────────────────────────────────┤
│                                                                  │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐  │
│  │  Execute │───▶│ Evaluate │───▶│ Diagnose │───▶│  Mutate  │  │
│  │   SOP    │    │  Metrics │    │ Weakness │    │   SOP    │  │
│  └──────────┘    └──────────┘    └──────────┘    └──────────┘  │
│       ▲                                                │         │
│       │                                                │         │
│       └────────────────────────────────────────────────┘         │
│                                                                  │
│  Mutation Types:                                                 │
│  • PROMPT_CHANGE  - Evolve prompts based on failure patterns    │
│  • PARAM_TWEAK    - Adjust model parameters (temp, tokens)      │
│  • ADD_ROLE       - Add new agents to the pipeline              │
│  • REMOVE_ROLE    - Simplify by removing agents                 │
│  • REWIRE_GRAPH   - Change how agents connect and communicate   │
│  • CROSSOVER      - Combine successful strategies               │
│                                                                  │
└─────────────────────────────────────────────────────────────────┘

Unlike prompt-only optimizers, SIARE can evolve the structure of your pipeline - adding, removing, and rewiring agents to find optimal architectures.

Using MAP-Elites, SIARE maintains a diverse population of high-performing solutions. You don't just get one answer - you get a range of options trading off different metrics.

Three evolution modes to match your needs:

• Single-turn: Fast, cheap mutations via Diagnostician + Architect (default)
• Agentic: Multi-turn variation with tool access — the agent diagnoses, proposes, validates, and iterates
• Adaptive: Starts single-turn, automatically escalates to agentic when stagnation is detected

When evolution stalls, a SupervisorAgent analyzes the trajectory and redirects exploration toward unexplored territory.

Enterprise features integrate cleanly via hooks without modifying core code:

• Audit logging
• Usage billing
• Approval gates
• Custom metrics

Evaluate your pipelines against standard datasets:

• HotpotQA (multi-hop reasoning)
• Natural Questions
• Custom evaluation suites

# Core package
pip install siare

# With LLM providers
pip install siare[llm]

# With embeddings support
pip install siare[embeddings]

# Everything
pip install siare[full]

from siare import ProcessConfig, Role, GraphEdge
from siare.services import DirectorService, ExecutionEngine, GenePool

# Define your pipeline
config = ProcessConfig(
    name="customer-support",
    version="1.0.0",
    roles=[
        Role(
            name="retriever",
            model="gpt-4o-mini",
            system_prompt="Find relevant documents...",
            tools=["vector_search"],
        ),
        Role(
            name="answerer",
            model="gpt-4o-mini",
            system_prompt="Answer based on retrieved context...",
        ),
    ],
    graph=[
        GraphEdge(source="retriever", target="answerer"),
    ],
)

# Run evolution
director = DirectorService(llm_provider)
gene_pool = GenePool()

for generation in range(10):
    # Execute and evaluate
    trace = await engine.execute(config, task)
    evaluation = await evaluator.evaluate(trace)

    # Diagnose and mutate
    diagnosis = await director.diagnose(evaluation)
    mutated = await director.mutate_sop(config, diagnosis)

    # Track in gene pool
    gene_pool.add(mutated, evaluation)

# Get best solution
best = gene_pool.get_pareto_frontier()[0]

• Customer Support - Simple Q&A over documents
• Clinical Trials - Complex multi-agent research assistant

• Quickstart Guide
• Configuration
• System Architecture
• Data Models
• Guides

We welcome contributions! See CONTRIBUTING.md for guidelines.

MIT License - see LICENSE for details.

Built with ❤️ by Synapti.ai