feat(ai): add CASTS for GeaFlow reasoning ability #737
Conversation
… Improved Decision Validation
…or structural signatures
…ss multiple files
…ality control Add native Gremlin simplePath() support to prevent pathological cycles in graph traversals. The implementation uses LLM-guided decision-making and AIMD confidence penalties rather than hard-coded restrictions, staying true to the system's learning philosophy. Key changes: - Add simplePath() step to Gremlin state machine for V, E, and P states - Implement per-request path history tracking in TraversalExecutor - Add cycle detection with configurable threshold and penalty modes - Enhance LLM Oracle prompts to recommend simplePath() for exploration goals - Add recent decision history context to improve LLM decision quality - Update configuration with CYCLE_PENALTY and CYCLE_DETECTION_THRESHOLD settings - Document design rationale and rejected alternatives in architecture.md - Add test case for simple path traversal validation
…ulations - Updated MetricsCollector to use Optional types for match_type, parent_node, parent_step_index, edge_label, sku_id, and decision parameters. - Enhanced EVALUATOR documentation to clarify evaluation phases and scoring mechanisms, including coverage rewards and penalties for cache misses. - Modified test cases in test_execution_lifecycle.py to align with new metrics structure and added tests for simple path execution. - Improved test coverage in test_gremlin_step_state_machine.py and test_lifecycle_integration.py to validate state transitions and integration with Gremlin state machine. - Refined threshold calculation tests to ensure monotonicity and boundary conditions. - Added dynamic execution environment constraints in documentation to clarify step legality in relation to current state and schema.
| # EMBEDDING SERVICE CONFIGURATION | ||
| # ============================================ | ||
| EMBEDDING_ENDPOINT = os.environ.get("EMBEDDING_ENDPOINT", "") | ||
| EMBEDDING_APIKEY = os.environ.get("EMBEDDING_APIKEY", "YOUR_EMBEDDING_API_KEY_HERE") |
There was a problem hiding this comment.
It's recommended to remove default values and provide explicit error messages when required values are empty.
| @@ -0,0 +1,210 @@ | |||
| """Configuration management for CASTS system. | |||
There was a problem hiding this comment.
It's better not to mix with Java code; place CASTS in a dedicated folder, such as geaflow-ai/casts, and why does casts have two nested layers with the same name?
There was a problem hiding this comment.
All files in the project must start with an Apache License header, even if they are blank.
There was a problem hiding this comment.
It is recommended to add a README in the root directory that briefly explains the origin of the CASTS module name, design goals, functional division of major components, code execution examples, or a demo, etc.
|
|
||
| @staticmethod | ||
| def get_state_and_options( | ||
| structural_signature: str, graph_schema: GraphSchema, node_id: str |
There was a problem hiding this comment.
Does graph_schema support dynamic modifications? The graph_schema is passed in each time the state machine is invoked, but there may be differences.
| pass | ||
|
|
||
| @abstractmethod | ||
| def get_valid_outgoing_edge_labels(self, node_id: str) -> list[str]: |
There was a problem hiding this comment.
In the interface for obtaining outgoing edge types, should we pass in the vertex label instead of the specific entity ID to confine the computation to the metadata level? Passing an ID would involve the process of fetching the entity → obtaining the label → computing the types of adjacent edges.
|
|
||
| @property | ||
| @abstractmethod | ||
| def goal_weights(self) -> list[int]: |
There was a problem hiding this comment.
Is it somewhat limiting to restrict weights to numeric types? Or, should they rather be float numbers?
|
|
||
| for source_id, out_edges in self._edges.items(): | ||
| if source_id in self._nodes: | ||
| out_labels = sorted({edge["label"] for edge in out_edges}) |
There was a problem hiding this comment.
Key constants should be declared as constants, including the target parameter below.
|
|
||
| def cleanup_low_confidence_skus(self) -> None: | ||
| """Remove SKUs that have fallen below the minimum confidence threshold.""" | ||
| self.knowledge_base = [sku for sku in self.knowledge_base if sku.confidence_score >= 0.1] |
There was a problem hiding this comment.
The hardcoded threshold of 0.1 is inconsistent with the min_confidence_threshold defined in config.py
2 participants
Summary
How was this PR tested?