Files
OpenJarvis/docs/architecture/learning.md
T
Jon Saad-FalconandClaude Opus 4.6 852259f18b Restructure codebase into 5-pillar architecture with MCP tool management, composition layer, and structured learning
Phase 1: Move RoutingContext to core/types.py, add RouterPolicy and QueryAnalyzer ABCs to intelligence/_stubs.py
Phase 2: Move memory backends to tools/storage/, convert memory/ to backward-compat shims
Phase 3: Add MCPToolAdapter, storage MCP tools, upgrade MCP server to spec 2025-11-25
Phase 4: Add SystemBuilder + JarvisSystem composition layer (system.py)
Phase 5: Add InstrumentedEngine for opt-in telemetry, simplify all agents
Phase 6: Add LearningPolicy ABC taxonomy with SFTPolicy, AgentAdvisorPolicy, ICLUpdaterPolicy
Phase 7: Update config schema (ToolsConfig, MCPConfig, TracesConfig, per-pillar learning policies)

Also: update all docs, README (DSPy-inspired), CLAUDE.md, and add logo assets.
1391 tests pass, 32 skipped. Zero new lint errors. Full backward compatibility via shims.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-22 05:57:13 +00:00

16 KiB

Learning & Traces

The Learning system is a cross-cutting concern that connects all four pillars through trace-driven feedback. It determines which model handles each query (router policies), records the full interaction as a trace, analyzes outcomes, and updates policies based on what worked.


LearningPolicy ABC Taxonomy

The learning system defines a hierarchy of learning policy ABCs. The base LearningPolicy ABC is specialized into three sub-ABCs corresponding to the three learnable concerns:

ABC Concern Description
IntelligenceLearningPolicy Model routing Determines which model handles a query (replaces the legacy RouterPolicy)
AgentLearningPolicy Agent behavior Advises on agent strategy (e.g., tool selection, turn limits)
ToolLearningPolicy Tool selection Recommends tool configurations based on query characteristics

All learning policies are registered in the LearningRegistry (in core/registry.py).

RouterPolicy ABC

The RouterPolicy ABC and the QueryAnalyzer ABC are defined in intelligence/_stubs.py:

# intelligence/_stubs.py
class RouterPolicy(ABC):
    @abstractmethod
    def select_model(self, context: RoutingContext) -> str:
        """Return the model registry key best suited for *context*."""

class QueryAnalyzer(ABC):
    @abstractmethod
    def analyze(self, query: str) -> RoutingContext:
        """Analyze a raw query string and return a RoutingContext."""

!!! note "Backward compatibility" The RouterPolicy and RoutingContext names are still importable from openjarvis.learning._stubs via backward-compatibility re-exports, but the canonical locations are now openjarvis.intelligence._stubs (for RouterPolicy and QueryAnalyzer) and openjarvis.core.types (for RoutingContext).

RoutingContext

The RoutingContext dataclass is now defined in core/types.py (moved from learning/_stubs.py):

# core/types.py
@dataclass(slots=True)
class RoutingContext:
    query: str = ""            # The raw query text
    query_length: int = 0      # Character count
    has_code: bool = False     # Whether code patterns were detected
    has_math: bool = False     # Whether math keywords were detected
    language: str = "en"       # Detected language
    urgency: float = 0.5      # 0 = low priority, 1 = real-time
    metadata: Dict[str, Any] = field(default_factory=dict)

RouterPolicyRegistry & LearningRegistry

Router policies are registered in the RouterPolicyRegistry and selected at runtime. Additionally, the LearningRegistry (in core/registry.py) manages the broader set of learning policies across the taxonomy.

The system ships with these router policies:

Registry Key Policy Class Status Description
heuristic HeuristicRouter Active Rule-based routing with 6 priority rules
learned TraceDrivenPolicy Active Learns from trace outcomes
grpo GRPORouterPolicy Stub Placeholder for future RL training
sft SFTPolicy Active Supervised fine-tuning policy (learns from labeled traces)

And these additional learning policies (registered in LearningRegistry):

Registry Key Policy Class Taxonomy Description
agent_advisor AgentAdvisorPolicy AgentLearningPolicy Advises on agent strategy based on trace patterns
icl_updater ICLUpdaterPolicy ToolLearningPolicy In-context learning updater for tool selection

Users select a policy via config.toml or the --router CLI flag:

[learning]
default_policy = "heuristic"
jarvis ask --router learned "What is the capital of France?"

The ensure_registered() Pattern

Learning modules use a lazy registration pattern to survive registry clearing in tests:

def ensure_registered() -> None:
    """Register TraceDrivenPolicy if not already present."""
    if not RouterPolicyRegistry.contains("learned"):
        RouterPolicyRegistry.register_value("learned", TraceDrivenPolicy)

ensure_registered()  # Called at module import time

This ensures that policies are available even after RouterPolicyRegistry.clear() is called in test teardown, because re-importing the module re-registers them.


HeuristicRouter (Heuristic Policy)

The HeuristicRouter is the default routing policy. It uses static rules to select models based on query characteristics. See the Intelligence Pillar documentation for full details on its six priority rules.

The heuristic_policy.py module wires the existing HeuristicRouter (from the Intelligence pillar) into the RouterPolicyRegistry:

# learning/heuristic_policy.py
def ensure_registered() -> None:
    if not RouterPolicyRegistry.contains("heuristic"):
        RouterPolicyRegistry.register_value("heuristic", HeuristicRouter)

ensure_registered()

TraceDrivenPolicy (Learned Policy)

The TraceDrivenPolicy learns from historical traces to determine which model performs best for different types of queries. Unlike the heuristic router's static rules, this policy adapts based on actual outcomes.

Query Classification

Queries are classified into broad categories for grouping:

Category Condition
code Contains code patterns (backticks, def, class, import, function)
math Contains math keywords (solve, integral, equation, calculate, compute)
short Query length < 50 characters
long Query length > 500 characters
general None of the above

Model Selection

When select_model() is called:

  1. Classify the query into a category
  2. If the policy map has an entry for this category and the confidence (sample count) exceeds min_samples (default: 5), use the learned model
  3. Otherwise, fall back to: default_model -> fallback_model -> first available model

Batch Updates via update_from_traces()

The primary update mechanism reads all traces from a TraceAnalyzer and recomputes the policy map:

from openjarvis.learning.trace_policy import TraceDrivenPolicy
from openjarvis.traces.analyzer import TraceAnalyzer
from openjarvis.traces.store import TraceStore

store = TraceStore("traces.db")
analyzer = TraceAnalyzer(store)
policy = TraceDrivenPolicy(
    analyzer=analyzer,
    available_models=["qwen3:8b", "llama3.2:3b", "deepseek-coder-v2:16b"],
    default_model="qwen3:8b",
)

# Recompute routing decisions from trace history
result = policy.update_from_traces()
# {"updated": True, "query_classes": 3, "total_traces": 150, "changes": {...}}

The update algorithm:

  1. Fetches all traces (optionally filtered by time range)
  2. Groups traces by query classification
  3. For each query class, scores each model using a composite score:
    • 60% success rate (fraction of traces with outcome="success")
    • 40% average feedback score (user quality ratings)
  4. Selects the model with the highest composite score for each query class
  5. Returns a summary of changes

Online Updates via observe()

For real-time policy updates after every interaction:

policy.observe(
    query="Write a Python function",
    model="deepseek-coder-v2:16b",
    outcome="success",
    feedback=0.9,
)

The online update uses a conservative strategy: it only switches the preferred model for a query class when the new model shows clearly better outcomes (feedback > 0.7) and the existing policy has fewer than min_samples observations.


SFTPolicy (Supervised Fine-Tuning)

The SFTPolicy (in learning/sft_policy.py) is an IntelligenceLearningPolicy that learns routing decisions from labeled trace data using supervised fine-tuning principles. Unlike TraceDrivenPolicy which uses online aggregation, SFTPolicy trains a mapping from query features to model choices based on curated, high-quality trace examples.

from openjarvis.learning.sft_policy import SFTPolicy

AgentAdvisorPolicy

The AgentAdvisorPolicy (in learning/agent_advisor.py) is an AgentLearningPolicy that advises on agent strategy -- for example, recommending tool sets, turn limits, or agent type -- based on patterns observed in historical traces.

from openjarvis.learning.agent_advisor import AgentAdvisorPolicy

ICLUpdaterPolicy

The ICLUpdaterPolicy (in learning/icl_updater.py) is a ToolLearningPolicy that uses in-context learning to update tool selection and configuration. It analyzes recent tool-call traces to recommend which tools should be enabled for different query types.

from openjarvis.learning.icl_updater import ICLUpdaterPolicy

GRPORouterPolicy (Stub)

The GRPORouterPolicy is a placeholder for future reinforcement learning-based routing. Currently, calling select_model() raises NotImplementedError:

class GRPORouterPolicy(RouterPolicy):
    def select_model(self, context: RoutingContext) -> str:
        raise NotImplementedError(
            "GRPORouterPolicy is not yet implemented. "
            "GRPO training will be available in a future phase."
        )

RewardFunction ABC

The RewardFunction ABC defines how to score completed inferences for use in training:

class RewardFunction(ABC):
    @abstractmethod
    def compute(
        self,
        context: RoutingContext,
        model_key: str,
        response: str,
        **kwargs: Any,
    ) -> float:
        """Return a reward in [0, 1]."""

HeuristicRewardFunction

The built-in reward function computes a weighted combination of three factors:

Factor Weight (default) Normalization Score Range
Latency 0.4 1 - (latency / max_latency) 0 = 30s+, 1 = instant
Cost 0.3 1 - (cost / max_cost) 0 = $0.01+, 1 = free
Efficiency 0.3 completion_tokens / total_tokens 0 = all prompt, 1 = all completion
from openjarvis.learning.heuristic_reward import HeuristicRewardFunction

reward_fn = HeuristicRewardFunction(
    weight_latency=0.4,
    weight_cost=0.3,
    weight_efficiency=0.3,
    max_latency=30.0,   # seconds
    max_cost=0.01,       # USD
)

reward = reward_fn.compute(
    context=routing_context,
    model_key="qwen3:8b",
    response="The answer is 42.",
    latency_seconds=1.2,
    cost_usd=0.0,
    prompt_tokens=50,
    completion_tokens=10,
)
# Returns a float in [0, 1]

Trace System

The trace system records the full sequence of steps in every agent interaction, providing the raw data that the learning system uses to improve.

TraceStore

TraceStore is an append-only SQLite store for interaction traces:

from openjarvis.traces.store import TraceStore

store = TraceStore("~/.openjarvis/traces.db")
store.save(trace)                          # Persist a complete trace
trace = store.get("abc123")                # Retrieve by trace ID
traces = store.list_traces(                # Query with filters
    agent="orchestrator",
    model="qwen3:8b",
    outcome="success",
    since=1700000000.0,
    limit=100,
)
count = store.count()                      # Total trace count

Database schema:

  • traces table -- one row per interaction (trace_id, query, agent, model, engine, result, outcome, feedback, timing, tokens, metadata)
  • trace_steps table -- one row per step within a trace (step_type, timestamp, duration, input, output, metadata)

EventBus integration: The store can subscribe to TRACE_COMPLETE events for automatic persistence:

store.subscribe_to_bus(bus)
# Any TRACE_COMPLETE event will now auto-save the trace

TraceCollector

TraceCollector wraps any BaseAgent and automatically records a Trace for every run() call:

from openjarvis.traces.collector import TraceCollector

agent = OrchestratorAgent(engine, model, tools=tools, bus=bus)
collector = TraceCollector(agent, store=trace_store, bus=bus)

result = collector.run("What is 2+2?")
# Trace is automatically saved to trace_store

How it works:

  1. Subscribes to EventBus events before running the agent:
    • INFERENCE_START / INFERENCE_END -- creates GENERATE steps
    • TOOL_CALL_START / TOOL_CALL_END -- creates TOOL_CALL steps
    • MEMORY_RETRIEVE -- creates RETRIEVE steps
  2. Runs the wrapped agent's run() method
  3. Unsubscribes from events
  4. Adds a final RESPOND step
  5. Builds a Trace object with all collected steps
  6. Saves to the TraceStore and publishes TRACE_COMPLETE

TraceAnalyzer

TraceAnalyzer provides a read-only query layer over stored traces, computing aggregated statistics:

from openjarvis.traces.analyzer import TraceAnalyzer

analyzer = TraceAnalyzer(store)

# Overall summary
summary = analyzer.summary()
# TraceSummary(total_traces=150, avg_latency=2.3, success_rate=0.85, ...)

# Stats grouped by (model, agent) routing decisions
route_stats = analyzer.per_route_stats()
# [RouteStats(model="qwen3:8b", agent="orchestrator", count=45, avg_latency=1.8, ...), ...]

# Stats grouped by tool
tool_stats = analyzer.per_tool_stats()
# [ToolStats(tool_name="calculator", call_count=23, avg_latency=0.01, success_rate=1.0), ...]

# Find traces matching query characteristics
code_traces = analyzer.traces_for_query_type(has_code=True)

# Export traces as plain dicts (for JSON serialization)
exported = analyzer.export_traces(limit=1000)

Computed statistics:

Dataclass Fields
TraceSummary total_traces, total_steps, avg_steps_per_trace, avg_latency, avg_tokens, success_rate, step_type_distribution
RouteStats model, agent, count, avg_latency, avg_tokens, success_rate, avg_feedback
ToolStats tool_name, call_count, avg_latency, success_rate

The Learning Loop

The trace-driven learning loop connects all the pieces:

graph TB
    subgraph "Runtime"
        Q["User Query"] --> AGT["Agent executes"]
        AGT --> ENG["Engine generates"]
        ENG --> RESP["Response returned"]
    end

    subgraph "Recording"
        AGT -.->|"events"| COL["TraceCollector"]
        ENG -.->|"events"| COL
        COL -->|"save"| STO["TraceStore<br/>(SQLite)"]
    end

    subgraph "Analysis"
        STO -->|"read"| ANA["TraceAnalyzer"]
        ANA -->|"summary(),<br/>per_route_stats()"| STATS["Aggregated<br/>Statistics"]
    end

    subgraph "Learning"
        STATS -->|"update_from_traces()"| POL["TraceDrivenPolicy"]
        POL -->|"select_model()"| Q
    end

    style Q fill:#e1f5fe
    style RESP fill:#e8f5e9
    style POL fill:#fff3e0

Step-by-step cycle:

  1. Query arrives -- The system needs to select a model
  2. Router policy selects model -- TraceDrivenPolicy.select_model() checks the learned policy map; falls back to heuristic if insufficient data
  3. Agent executes -- The agent processes the query, calling tools and memory as needed
  4. Events captured -- The TraceCollector captures all events (inference, tool calls, memory retrieval) during execution
  5. Trace saved -- A complete Trace with all TraceStep objects is saved to TraceStore
  6. Analysis -- Periodically, TraceAnalyzer computes aggregate statistics from stored traces
  7. Policy update -- TraceDrivenPolicy.update_from_traces() recomputes the query_class -> model mapping based on success rates and feedback scores
  8. Better routing -- The next query benefits from the updated routing decisions

Trace Data Model

Each interaction produces a Trace containing multiple TraceStep objects:

Trace
  trace_id: "a1b2c3d4e5f6"
  query: "What is 2+2?"
  agent: "orchestrator"
  model: "qwen3:8b"
  engine: "ollama"
  steps:
    [0] GENERATE  -- model inference, 0.8s, 150 tokens
    [1] TOOL_CALL -- calculator, 0.01s, success
    [2] GENERATE  -- model inference, 0.5s, 80 tokens
    [3] RESPOND   -- final answer
  result: "2+2 = 4"
  outcome: "success"
  feedback: 1.0
  total_latency_seconds: 1.31
  total_tokens: 230

Step types:

StepType Description Created By
ROUTE Model selection decision Router policy
RETRIEVE Memory search Memory backend
GENERATE LLM inference call Engine
TOOL_CALL Tool execution ToolExecutor
RESPOND Final response TraceCollector