mirror of
https://github.com/trustgraph-ai/trustgraph.git
synced 2026-04-25 00:16:23 +02:00
master -> 1.5 (README updates) (#552)
This commit is contained in:
cybermaggedon
GitHub
parent
ad35656811
commit
51107008fd
18 changed files with 891 additions and 120 deletions
106
docs/tech-specs/ARCHITECTURE_PRINCIPLES.md
Normal file
106
docs/tech-specs/ARCHITECTURE_PRINCIPLES.md
Normal file
|
|
@ -0,0 +1,106 @@
|
|||
# Knowledge Graph Architecture Foundations
|
||||
|
||||
## Foundation 1: Subject-Predicate-Object (SPO) Graph Model
|
||||
**Decision**: Adopt SPO/RDF as the core knowledge representation model
|
||||
|
||||
**Rationale**:
|
||||
- Provides maximum flexibility and interoperability with existing graph technologies
|
||||
- Enables seamless translation to other graph query languages (e.g., SPO → Cypher, but not vice versa)
|
||||
- Creates a foundation that "unlocks a lot" of downstream capabilities
|
||||
- Supports both node-to-node relationships (SPO) and node-to-literal relationships (RDF)
|
||||
|
||||
**Implementation**:
|
||||
- Core data structure: `node → edge → {node | literal}`
|
||||
- Maintain compatibility with RDF standards while supporting extended SPO operations
|
||||
|
||||
## Foundation 2: LLM-Native Knowledge Graph Integration
|
||||
**Decision**: Optimize knowledge graph structure and operations for LLM interaction
|
||||
|
||||
**Rationale**:
|
||||
- Primary use case involves LLMs interfacing with knowledge graphs
|
||||
- Graph technology choices must prioritize LLM compatibility over other considerations
|
||||
- Enables natural language processing workflows that leverage structured knowledge
|
||||
|
||||
**Implementation**:
|
||||
- Design graph schemas that LLMs can effectively reason about
|
||||
- Optimize for common LLM interaction patterns
|
||||
|
||||
## Foundation 3: Embedding-Based Graph Navigation
|
||||
**Decision**: Implement direct mapping from natural language queries to graph nodes via embeddings
|
||||
|
||||
**Rationale**:
|
||||
- Enables the simplest possible path from NLP query to graph navigation
|
||||
- Avoids complex intermediate query generation steps
|
||||
- Provides efficient semantic search capabilities within the graph structure
|
||||
|
||||
**Implementation**:
|
||||
- `NLP Query → Graph Embeddings → Graph Nodes`
|
||||
- Maintain embedding representations for all graph entities
|
||||
- Support direct semantic similarity matching for query resolution
|
||||
|
||||
## Foundation 4: Distributed Entity Resolution with Deterministic Identifiers
|
||||
**Decision**: Support parallel knowledge extraction with deterministic entity identification (80% rule)
|
||||
|
||||
**Rationale**:
|
||||
- **Ideal**: Single-process extraction with complete state visibility enables perfect entity resolution
|
||||
- **Reality**: Scalability requirements demand parallel processing capabilities
|
||||
- **Compromise**: Design for deterministic entity identification across distributed processes
|
||||
|
||||
**Implementation**:
|
||||
- Develop mechanisms for generating consistent, unique identifiers across different knowledge extractors
|
||||
- Same entity mentioned in different processes must resolve to the same identifier
|
||||
- Acknowledge that ~20% of edge cases may require alternative processing models
|
||||
- Design fallback mechanisms for complex entity resolution scenarios
|
||||
|
||||
## Foundation 5: Event-Driven Architecture with Publish-Subscribe
|
||||
**Decision**: Implement pub-sub messaging system for system coordination
|
||||
|
||||
**Rationale**:
|
||||
- Enables loose coupling between knowledge extraction, storage, and query components
|
||||
- Supports real-time updates and notifications across the system
|
||||
- Facilitates scalable, distributed processing workflows
|
||||
|
||||
**Implementation**:
|
||||
- Message-driven coordination between system components
|
||||
- Event streams for knowledge updates, extraction completion, and query results
|
||||
|
||||
## Foundation 6: Reentrant Agent Communication
|
||||
**Decision**: Support reentrant pub-sub operations for agent-based processing
|
||||
|
||||
**Rationale**:
|
||||
- Enables sophisticated agent workflows where agents can trigger and respond to each other
|
||||
- Supports complex, multi-step knowledge processing pipelines
|
||||
- Allows for recursive and iterative processing patterns
|
||||
|
||||
**Implementation**:
|
||||
- Pub-sub system must handle reentrant calls safely
|
||||
- Agent coordination mechanisms that prevent infinite loops
|
||||
- Support for agent workflow orchestration
|
||||
|
||||
## Foundation 7: Columnar Data Store Integration
|
||||
**Decision**: Ensure query compatibility with columnar storage systems
|
||||
|
||||
**Rationale**:
|
||||
- Enables efficient analytical queries over large knowledge datasets
|
||||
- Supports business intelligence and reporting use cases
|
||||
- Bridges graph-based knowledge representation with traditional analytical workflows
|
||||
|
||||
**Implementation**:
|
||||
- Query translation layer: Graph queries → Columnar queries
|
||||
- Hybrid storage strategy supporting both graph operations and analytical workloads
|
||||
- Maintain query performance across both paradigms
|
||||
|
||||
---
|
||||
|
||||
## Architecture Principles Summary
|
||||
|
||||
1. **Flexibility First**: SPO/RDF model provides maximum adaptability
|
||||
2. **LLM Optimization**: All design decisions consider LLM interaction requirements
|
||||
3. **Semantic Efficiency**: Direct embedding-to-node mapping for optimal query performance
|
||||
4. **Pragmatic Scalability**: Balance perfect accuracy with practical distributed processing
|
||||
5. **Event-Driven Coordination**: Pub-sub enables loose coupling and scalability
|
||||
6. **Agent-Friendly**: Support complex, multi-agent processing workflows
|
||||
7. **Analytical Compatibility**: Bridge graph and columnar paradigms for comprehensive querying
|
||||
|
||||
These foundations establish a knowledge graph architecture that balances theoretical rigor with practical scalability requirements, optimized for LLM integration and distributed processing.
|
||||
|
||||
169
docs/tech-specs/LOGGING_STRATEGY.md
Normal file
169
docs/tech-specs/LOGGING_STRATEGY.md
Normal file
|
|
@ -0,0 +1,169 @@
|
|||
# TrustGraph Logging Strategy
|
||||
|
||||
## Overview
|
||||
|
||||
TrustGraph uses Python's built-in `logging` module for all logging operations. This provides a standardized, flexible approach to logging across all components of the system.
|
||||
|
||||
## Default Configuration
|
||||
|
||||
### Logging Level
|
||||
- **Default Level**: `INFO`
|
||||
- **Debug Mode**: `DEBUG` (enabled via command-line argument)
|
||||
- **Production**: `WARNING` or `ERROR` as appropriate
|
||||
|
||||
### Output Destination
|
||||
All logs should be written to **standard output (stdout)** to ensure compatibility with containerized environments and log aggregation systems.
|
||||
|
||||
## Implementation Guidelines
|
||||
|
||||
### 1. Logger Initialization
|
||||
|
||||
Each module should create its own logger using the module's `__name__`:
|
||||
|
||||
```python
|
||||
import logging
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
```
|
||||
|
||||
### 2. Centralized Configuration
|
||||
|
||||
The logging configuration should be centralized in `async_processor.py` (or a dedicated logging configuration module) since it's inherited by much of the codebase:
|
||||
|
||||
```python
|
||||
import logging
|
||||
import argparse
|
||||
|
||||
def setup_logging(log_level='INFO'):
|
||||
"""Configure logging for the entire application"""
|
||||
logging.basicConfig(
|
||||
level=getattr(logging, log_level.upper()),
|
||||
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s',
|
||||
handlers=[logging.StreamHandler()]
|
||||
)
|
||||
|
||||
def parse_args():
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
'--log-level',
|
||||
default='INFO',
|
||||
choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'],
|
||||
help='Set the logging level (default: INFO)'
|
||||
)
|
||||
return parser.parse_args()
|
||||
|
||||
# In main execution
|
||||
if __name__ == '__main__':
|
||||
args = parse_args()
|
||||
setup_logging(args.log_level)
|
||||
```
|
||||
|
||||
### 3. Logging Best Practices
|
||||
|
||||
#### Log Levels Usage
|
||||
- **DEBUG**: Detailed information for diagnosing problems (variable values, function entry/exit)
|
||||
- **INFO**: General informational messages (service started, configuration loaded, processing milestones)
|
||||
- **WARNING**: Warning messages for potentially harmful situations (deprecated features, recoverable errors)
|
||||
- **ERROR**: Error messages for serious problems (failed operations, exceptions)
|
||||
- **CRITICAL**: Critical messages for system failures requiring immediate attention
|
||||
|
||||
#### Message Format
|
||||
```python
|
||||
# Good - includes context
|
||||
logger.info(f"Processing document: {doc_id}, size: {doc_size} bytes")
|
||||
logger.error(f"Failed to connect to database: {error}", exc_info=True)
|
||||
|
||||
# Avoid - lacks context
|
||||
logger.info("Processing document")
|
||||
logger.error("Connection failed")
|
||||
```
|
||||
|
||||
#### Performance Considerations
|
||||
```python
|
||||
# Use lazy formatting for expensive operations
|
||||
logger.debug("Expensive operation result: %s", expensive_function())
|
||||
|
||||
# Check log level for very expensive debug operations
|
||||
if logger.isEnabledFor(logging.DEBUG):
|
||||
debug_data = compute_expensive_debug_info()
|
||||
logger.debug(f"Debug data: {debug_data}")
|
||||
```
|
||||
|
||||
### 4. Structured Logging
|
||||
|
||||
For complex data, use structured logging:
|
||||
|
||||
```python
|
||||
logger.info("Request processed", extra={
|
||||
'request_id': request_id,
|
||||
'duration_ms': duration,
|
||||
'status_code': status_code,
|
||||
'user_id': user_id
|
||||
})
|
||||
```
|
||||
|
||||
### 5. Exception Logging
|
||||
|
||||
Always include stack traces for exceptions:
|
||||
|
||||
```python
|
||||
try:
|
||||
process_data()
|
||||
except Exception as e:
|
||||
logger.error(f"Failed to process data: {e}", exc_info=True)
|
||||
raise
|
||||
```
|
||||
|
||||
### 6. Async Logging Considerations
|
||||
|
||||
For async code, ensure thread-safe logging:
|
||||
|
||||
```python
|
||||
import asyncio
|
||||
import logging
|
||||
|
||||
async def async_operation():
|
||||
logger = logging.getLogger(__name__)
|
||||
logger.info(f"Starting async operation in task: {asyncio.current_task().get_name()}")
|
||||
```
|
||||
|
||||
## Environment Variables
|
||||
|
||||
Support environment-based configuration as a fallback:
|
||||
|
||||
```python
|
||||
import os
|
||||
|
||||
log_level = os.environ.get('TRUSTGRAPH_LOG_LEVEL', 'INFO')
|
||||
```
|
||||
|
||||
## Testing
|
||||
|
||||
During tests, consider using a different logging configuration:
|
||||
|
||||
```python
|
||||
# In test setup
|
||||
logging.getLogger().setLevel(logging.WARNING) # Reduce noise during tests
|
||||
```
|
||||
|
||||
## Monitoring Integration
|
||||
|
||||
Ensure log format is compatible with monitoring tools:
|
||||
- Include timestamps in ISO format
|
||||
- Use consistent field names
|
||||
- Include correlation IDs where applicable
|
||||
- Structure logs for easy parsing (JSON format for production)
|
||||
|
||||
## Security Considerations
|
||||
|
||||
- Never log sensitive information (passwords, API keys, personal data)
|
||||
- Sanitize user input before logging
|
||||
- Use placeholders for sensitive fields: `user_id=****1234`
|
||||
|
||||
## Migration Path
|
||||
|
||||
For existing code using print statements:
|
||||
1. Replace `print()` with appropriate logger calls
|
||||
2. Choose appropriate log levels based on message importance
|
||||
3. Add context to make logs more useful
|
||||
4. Test logging output at different levels
|
||||
91
docs/tech-specs/SCHEMA_REFACTORING_PROPOSAL.md
Normal file
91
docs/tech-specs/SCHEMA_REFACTORING_PROPOSAL.md
Normal file
|
|
@ -0,0 +1,91 @@
|
|||
# Schema Directory Refactoring Proposal
|
||||
|
||||
## Current Issues
|
||||
|
||||
1. **Flat structure** - All schemas in one directory makes it hard to understand relationships
|
||||
2. **Mixed concerns** - Core types, domain objects, and API contracts all mixed together
|
||||
3. **Unclear naming** - Files like "object.py", "types.py", "topic.py" don't clearly indicate their purpose
|
||||
4. **No clear layering** - Can't easily see what depends on what
|
||||
|
||||
## Proposed Structure
|
||||
|
||||
```
|
||||
trustgraph-base/trustgraph/schema/
|
||||
├── __init__.py
|
||||
├── core/ # Core primitive types used everywhere
|
||||
│ ├── __init__.py
|
||||
│ ├── primitives.py # Error, Value, Triple, Field, RowSchema
|
||||
│ ├── metadata.py # Metadata record
|
||||
│ └── topic.py # Topic utilities
|
||||
│
|
||||
├── knowledge/ # Knowledge domain models and extraction
|
||||
│ ├── __init__.py
|
||||
│ ├── graph.py # EntityContext, EntityEmbeddings, Triples
|
||||
│ ├── document.py # Document, TextDocument, Chunk
|
||||
│ ├── knowledge.py # Knowledge extraction types
|
||||
│ ├── embeddings.py # All embedding-related types (moved from multiple files)
|
||||
│ └── nlp.py # Definition, Topic, Relationship, Fact types
|
||||
│
|
||||
└── services/ # Service request/response contracts
|
||||
├── __init__.py
|
||||
├── llm.py # TextCompletion, Embeddings, Tool requests/responses
|
||||
├── retrieval.py # GraphRAG, DocumentRAG queries/responses
|
||||
├── query.py # GraphEmbeddingsRequest/Response, DocumentEmbeddingsRequest/Response
|
||||
├── agent.py # Agent requests/responses
|
||||
├── flow.py # Flow requests/responses
|
||||
├── prompt.py # Prompt service requests/responses
|
||||
├── config.py # Configuration service
|
||||
├── library.py # Librarian service
|
||||
└── lookup.py # Lookup service
|
||||
```
|
||||
|
||||
## Key Changes
|
||||
|
||||
1. **Hierarchical organization** - Clear separation between core types, knowledge models, and service contracts
|
||||
2. **Better naming**:
|
||||
- `types.py` → `core/primitives.py` (clearer purpose)
|
||||
- `object.py` → Split between appropriate files based on actual content
|
||||
- `documents.py` → `knowledge/document.py` (singular, consistent)
|
||||
- `models.py` → `services/llm.py` (clearer what kind of models)
|
||||
- `prompt.py` → Split: service parts to `services/prompt.py`, data types to `knowledge/nlp.py`
|
||||
|
||||
3. **Logical grouping**:
|
||||
- All embedding types consolidated in `knowledge/embeddings.py`
|
||||
- All LLM-related service contracts in `services/llm.py`
|
||||
- Clear separation of request/response pairs in services directory
|
||||
- Knowledge extraction types grouped with other knowledge domain models
|
||||
|
||||
4. **Dependency clarity**:
|
||||
- Core types have no dependencies
|
||||
- Knowledge models depend only on core
|
||||
- Service contracts can depend on both core and knowledge models
|
||||
|
||||
## Migration Benefits
|
||||
|
||||
1. **Easier navigation** - Developers can quickly find what they need
|
||||
2. **Better modularity** - Clear boundaries between different concerns
|
||||
3. **Simpler imports** - More intuitive import paths
|
||||
4. **Future-proof** - Easy to add new knowledge types or services without cluttering
|
||||
|
||||
## Example Import Changes
|
||||
|
||||
```python
|
||||
# Before
|
||||
from trustgraph.schema import Error, Triple, GraphEmbeddings, TextCompletionRequest
|
||||
|
||||
# After
|
||||
from trustgraph.schema.core import Error, Triple
|
||||
from trustgraph.schema.knowledge import GraphEmbeddings
|
||||
from trustgraph.schema.services import TextCompletionRequest
|
||||
```
|
||||
|
||||
## Implementation Notes
|
||||
|
||||
1. Keep backward compatibility by maintaining imports in root `__init__.py`
|
||||
2. Move files gradually, updating imports as needed
|
||||
3. Consider adding a `legacy.py` that imports everything for transition period
|
||||
4. Update documentation to reflect new structure
|
||||
|
||||
<function_calls>
|
||||
<invoke name="TodoWrite">
|
||||
<parameter name="todos">[{"id": "1", "content": "Examine current schema directory structure", "status": "completed", "priority": "high"}, {"id": "2", "content": "Analyze schema files and their purposes", "status": "completed", "priority": "high"}, {"id": "3", "content": "Propose improved naming and structure", "status": "completed", "priority": "high"}]
|
||||
253
docs/tech-specs/STRUCTURED_DATA.md
Normal file
253
docs/tech-specs/STRUCTURED_DATA.md
Normal file
|
|
@ -0,0 +1,253 @@
|
|||
# Structured Data Technical Specification
|
||||
|
||||
## Overview
|
||||
|
||||
This specification describes the integration of TrustGraph with structured data flows, enabling the system to work with data that can be represented as rows in tables or objects in object stores. The integration supports four primary use cases:
|
||||
|
||||
1. **Unstructured to Structured Extraction**: Read unstructured data sources, identify and extract object structures, and store them in a tabular format
|
||||
2. **Structured Data Ingestion**: Load data that is already in structured formats directly into the structured store alongside extracted data
|
||||
3. **Natural Language Querying**: Convert natural language questions into structured queries to extract matching data from the store
|
||||
4. **Direct Structured Querying**: Execute structured queries directly against the data store for precise data retrieval
|
||||
|
||||
## Goals
|
||||
|
||||
- **Unified Data Access**: Provide a single interface for accessing both structured and unstructured data within TrustGraph
|
||||
- **Seamless Integration**: Enable smooth interoperability between TrustGraph's graph-based knowledge representation and traditional structured data formats
|
||||
- **Flexible Extraction**: Support automatic extraction of structured data from various unstructured sources (documents, text, etc.)
|
||||
- **Query Versatility**: Allow users to query data using both natural language and structured query languages
|
||||
- **Data Consistency**: Maintain data integrity and consistency across different data representations
|
||||
- **Performance Optimization**: Ensure efficient storage and retrieval of structured data at scale
|
||||
- **Schema Flexibility**: Support both schema-on-write and schema-on-read approaches to accommodate diverse data sources
|
||||
- **Backwards Compatibility**: Preserve existing TrustGraph functionality while adding structured data capabilities
|
||||
|
||||
## Background
|
||||
|
||||
TrustGraph currently excels at processing unstructured data and building knowledge graphs from diverse sources. However, many enterprise use cases involve data that is inherently structured - customer records, transaction logs, inventory databases, and other tabular datasets. These structured datasets often need to be analyzed alongside unstructured content to provide comprehensive insights.
|
||||
|
||||
Current limitations include:
|
||||
- No native support for ingesting pre-structured data formats (CSV, JSON arrays, database exports)
|
||||
- Inability to preserve the inherent structure when extracting tabular data from documents
|
||||
- Lack of efficient querying mechanisms for structured data patterns
|
||||
- Missing bridge between SQL-like queries and TrustGraph's graph queries
|
||||
|
||||
This specification addresses these gaps by introducing a structured data layer that complements TrustGraph's existing capabilities. By supporting structured data natively, TrustGraph can:
|
||||
- Serve as a unified platform for both structured and unstructured data analysis
|
||||
- Enable hybrid queries that span both graph relationships and tabular data
|
||||
- Provide familiar interfaces for users accustomed to working with structured data
|
||||
- Unlock new use cases in data integration and business intelligence
|
||||
|
||||
## Technical Design
|
||||
|
||||
### Architecture
|
||||
|
||||
The structured data integration requires the following technical components:
|
||||
|
||||
1. **NLP-to-Structured-Query Service**
|
||||
- Converts natural language questions into structured queries
|
||||
- Supports multiple query language targets (initially SQL-like syntax)
|
||||
- Integrates with existing TrustGraph NLP capabilities
|
||||
|
||||
Module: trustgraph-flow/trustgraph/query/nlp_query/cassandra
|
||||
|
||||
2. **Configuration Schema Support** ✅ **[COMPLETE]**
|
||||
- Extended configuration system to store structured data schemas
|
||||
- Support for defining table structures, field types, and relationships
|
||||
- Schema versioning and migration capabilities
|
||||
|
||||
3. **Object Extraction Module** ✅ **[COMPLETE]**
|
||||
- Enhanced knowledge extractor flow integration
|
||||
- Identifies and extracts structured objects from unstructured sources
|
||||
- Maintains provenance and confidence scores
|
||||
- Registers a config handler (example: trustgraph-flow/trustgraph/prompt/template/service.py) to receive config data and decode schema information
|
||||
- Receives objects and decodes them to ExtractedObject objects for delivery on the Pulsar queue
|
||||
- NOTE: There's existing code at `trustgraph-flow/trustgraph/extract/object/row/`. This was a previous attempt and will need to be majorly refactored as it doesn't conform to current APIs. Use it if it's useful, start from scratch if not.
|
||||
- Requires a command-line interface: `kg-extract-objects`
|
||||
|
||||
Module: trustgraph-flow/trustgraph/extract/kg/objects/
|
||||
|
||||
4. **Structured Store Writer Module** ✅ **[COMPLETE]**
|
||||
- Receives objects in ExtractedObject format from Pulsar queues
|
||||
- Initial implementation targeting Apache Cassandra as the structured data store
|
||||
- Handles dynamic table creation based on schemas encountered
|
||||
- Manages schema-to-Cassandra table mapping and data transformation
|
||||
- Provides batch and streaming write operations for performance optimization
|
||||
- No Pulsar outputs - this is a terminal service in the data flow
|
||||
|
||||
**Schema Handling**:
|
||||
- Monitors incoming ExtractedObject messages for schema references
|
||||
- When a new schema is encountered for the first time, automatically creates the corresponding Cassandra table
|
||||
- Maintains a cache of known schemas to avoid redundant table creation attempts
|
||||
- Should consider whether to receive schema definitions directly or rely on schema names in ExtractedObject messages
|
||||
|
||||
**Cassandra Table Mapping**:
|
||||
- Keyspace is named after the `user` field from ExtractedObject's Metadata
|
||||
- Table is named after the `schema_name` field from ExtractedObject
|
||||
- Collection from Metadata becomes part of the partition key to ensure:
|
||||
- Natural data distribution across Cassandra nodes
|
||||
- Efficient queries within a specific collection
|
||||
- Logical isolation between different data imports/sources
|
||||
- Primary key structure: `PRIMARY KEY ((collection, <schema_primary_key_fields>), <clustering_keys>)`
|
||||
- Collection is always the first component of the partition key
|
||||
- Schema-defined primary key fields follow as part of the composite partition key
|
||||
- This requires queries to specify the collection, ensuring predictable performance
|
||||
- Field definitions map to Cassandra columns with type conversions:
|
||||
- `string` → `text`
|
||||
- `integer` → `int` or `bigint` based on size hint
|
||||
- `float` → `float` or `double` based on precision needs
|
||||
- `boolean` → `boolean`
|
||||
- `timestamp` → `timestamp`
|
||||
- `enum` → `text` with application-level validation
|
||||
- Indexed fields create Cassandra secondary indexes (excluding fields already in the primary key)
|
||||
- Required fields are enforced at the application level (Cassandra doesn't support NOT NULL)
|
||||
|
||||
**Object Storage**:
|
||||
- Extracts values from ExtractedObject.values map
|
||||
- Performs type conversion and validation before insertion
|
||||
- Handles missing optional fields gracefully
|
||||
- Maintains metadata about object provenance (source document, confidence scores)
|
||||
- Supports idempotent writes to handle message replay scenarios
|
||||
|
||||
**Implementation Notes**:
|
||||
- Existing code at `trustgraph-flow/trustgraph/storage/objects/cassandra/` is outdated and doesn't comply with current APIs
|
||||
- Should reference `trustgraph-flow/trustgraph/storage/triples/cassandra` as an example of a working storage processor
|
||||
- Needs evaluation of existing code for any reusable components before deciding to refactor or rewrite
|
||||
|
||||
Module: trustgraph-flow/trustgraph/storage/objects/cassandra
|
||||
|
||||
5. **Structured Query Service**
|
||||
- Accepts structured queries in defined formats
|
||||
- Executes queries against the structured store
|
||||
- Returns objects matching query criteria
|
||||
- Supports pagination and result filtering
|
||||
|
||||
Module: trustgraph-flow/trustgraph/query/objects/cassandra
|
||||
|
||||
6. **Agent Tool Integration**
|
||||
- New tool class for agent frameworks
|
||||
- Enables agents to query structured data stores
|
||||
- Provides natural language and structured query interfaces
|
||||
- Integrates with existing agent decision-making processes
|
||||
|
||||
7. **Structured Data Ingestion Service**
|
||||
- Accepts structured data in multiple formats (JSON, CSV, XML)
|
||||
- Parses and validates incoming data against defined schemas
|
||||
- Converts data into normalized object streams
|
||||
- Emits objects to appropriate message queues for processing
|
||||
- Supports bulk uploads and streaming ingestion
|
||||
|
||||
Module: trustgraph-flow/trustgraph/decoding/structured
|
||||
|
||||
8. **Object Embedding Service**
|
||||
- Generates vector embeddings for structured objects
|
||||
- Enables semantic search across structured data
|
||||
- Supports hybrid search combining structured queries with semantic similarity
|
||||
- Integrates with existing vector stores
|
||||
|
||||
Module: trustgraph-flow/trustgraph/embeddings/object_embeddings/qdrant
|
||||
|
||||
### Data Models
|
||||
|
||||
#### Schema Storage Mechanism
|
||||
|
||||
Schemas are stored in TrustGraph's configuration system using the following structure:
|
||||
|
||||
- **Type**: `schema` (fixed value for all structured data schemas)
|
||||
- **Key**: The unique name/identifier of the schema (e.g., `customer_records`, `transaction_log`)
|
||||
- **Value**: JSON schema definition containing the structure
|
||||
|
||||
Example configuration entry:
|
||||
```
|
||||
Type: schema
|
||||
Key: customer_records
|
||||
Value: {
|
||||
"name": "customer_records",
|
||||
"description": "Customer information table",
|
||||
"fields": [
|
||||
{
|
||||
"name": "customer_id",
|
||||
"type": "string",
|
||||
"primary_key": true
|
||||
},
|
||||
{
|
||||
"name": "name",
|
||||
"type": "string",
|
||||
"required": true
|
||||
},
|
||||
{
|
||||
"name": "email",
|
||||
"type": "string",
|
||||
"required": true
|
||||
},
|
||||
{
|
||||
"name": "registration_date",
|
||||
"type": "timestamp"
|
||||
},
|
||||
{
|
||||
"name": "status",
|
||||
"type": "string",
|
||||
"enum": ["active", "inactive", "suspended"]
|
||||
}
|
||||
],
|
||||
"indexes": ["email", "registration_date"]
|
||||
}
|
||||
```
|
||||
|
||||
This approach allows:
|
||||
- Dynamic schema definition without code changes
|
||||
- Easy schema updates and versioning
|
||||
- Consistent integration with existing TrustGraph configuration management
|
||||
- Support for multiple schemas within a single deployment
|
||||
|
||||
### APIs
|
||||
|
||||
New APIs:
|
||||
- Pulsar schemas for above types
|
||||
- Pulsar interfaces in new flows
|
||||
- Need a means to specify schema types in flows so that flows know which
|
||||
schema types to load
|
||||
- APIs added to gateway and rev-gateway
|
||||
|
||||
Modified APIs:
|
||||
- Knowledge extraction endpoints - Add structured object output option
|
||||
- Agent endpoints - Add structured data tool support
|
||||
|
||||
### Implementation Details
|
||||
|
||||
Following existing conventions - these are just new processing modules.
|
||||
Everything is in the trustgraph-flow packages except for schema items
|
||||
in trustgraph-base.
|
||||
|
||||
Need some UI work in the Workbench to be able to demo / pilot this
|
||||
capability.
|
||||
|
||||
## Security Considerations
|
||||
|
||||
No extra considerations.
|
||||
|
||||
## Performance Considerations
|
||||
|
||||
Some questions around using Cassandra queries and indexes so that queries
|
||||
don't slow down.
|
||||
|
||||
## Testing Strategy
|
||||
|
||||
Use existing test strategy, will build unit, contract and integration tests.
|
||||
|
||||
## Migration Plan
|
||||
|
||||
None.
|
||||
|
||||
## Timeline
|
||||
|
||||
Not specified.
|
||||
|
||||
## Open Questions
|
||||
|
||||
- Can this be made to work with other store types? We're aiming to use
|
||||
interfaces which make modules which work with one store applicable to
|
||||
other stores.
|
||||
|
||||
## References
|
||||
|
||||
n/a.
|
||||
|
||||
139
docs/tech-specs/STRUCTURED_DATA_SCHEMAS.md
Normal file
139
docs/tech-specs/STRUCTURED_DATA_SCHEMAS.md
Normal file
|
|
@ -0,0 +1,139 @@
|
|||
# Structured Data Pulsar Schema Changes
|
||||
|
||||
## Overview
|
||||
|
||||
Based on the STRUCTURED_DATA.md specification, this document proposes the necessary Pulsar schema additions and modifications to support structured data capabilities in TrustGraph.
|
||||
|
||||
## Required Schema Changes
|
||||
|
||||
### 1. Core Schema Enhancements
|
||||
|
||||
#### Enhanced Field Definition
|
||||
The existing `Field` class in `core/primitives.py` needs additional properties:
|
||||
|
||||
```python
|
||||
class Field(Record):
|
||||
name = String()
|
||||
type = String() # int, string, long, bool, float, double, timestamp
|
||||
size = Integer()
|
||||
primary = Boolean()
|
||||
description = String()
|
||||
# NEW FIELDS:
|
||||
required = Boolean() # Whether field is required
|
||||
enum_values = Array(String()) # For enum type fields
|
||||
indexed = Boolean() # Whether field should be indexed
|
||||
```
|
||||
|
||||
### 2. New Knowledge Schemas
|
||||
|
||||
#### 2.1 Structured Data Submission
|
||||
New file: `knowledge/structured.py`
|
||||
|
||||
```python
|
||||
from pulsar.schema import Record, String, Bytes, Map
|
||||
from ..core.metadata import Metadata
|
||||
|
||||
class StructuredDataSubmission(Record):
|
||||
metadata = Metadata()
|
||||
format = String() # "json", "csv", "xml"
|
||||
schema_name = String() # Reference to schema in config
|
||||
data = Bytes() # Raw data to ingest
|
||||
options = Map(String()) # Format-specific options
|
||||
```
|
||||
|
||||
### 3. New Service Schemas
|
||||
|
||||
#### 3.1 NLP to Structured Query Service
|
||||
New file: `services/nlp_query.py`
|
||||
|
||||
```python
|
||||
from pulsar.schema import Record, String, Array, Map, Integer, Double
|
||||
from ..core.primitives import Error
|
||||
|
||||
class NLPToStructuredQueryRequest(Record):
|
||||
natural_language_query = String()
|
||||
max_results = Integer()
|
||||
context_hints = Map(String()) # Optional context for query generation
|
||||
|
||||
class NLPToStructuredQueryResponse(Record):
|
||||
error = Error()
|
||||
graphql_query = String() # Generated GraphQL query
|
||||
variables = Map(String()) # GraphQL variables if any
|
||||
detected_schemas = Array(String()) # Which schemas the query targets
|
||||
confidence = Double()
|
||||
```
|
||||
|
||||
#### 3.2 Structured Query Service
|
||||
New file: `services/structured_query.py`
|
||||
|
||||
```python
|
||||
from pulsar.schema import Record, String, Map, Array
|
||||
from ..core.primitives import Error
|
||||
|
||||
class StructuredQueryRequest(Record):
|
||||
query = String() # GraphQL query
|
||||
variables = Map(String()) # GraphQL variables
|
||||
operation_name = String() # Optional operation name for multi-operation documents
|
||||
|
||||
class StructuredQueryResponse(Record):
|
||||
error = Error()
|
||||
data = String() # JSON-encoded GraphQL response data
|
||||
errors = Array(String()) # GraphQL errors if any
|
||||
```
|
||||
|
||||
#### 2.2 Object Extraction Output
|
||||
New file: `knowledge/object.py`
|
||||
|
||||
```python
|
||||
from pulsar.schema import Record, String, Map, Double
|
||||
from ..core.metadata import Metadata
|
||||
|
||||
class ExtractedObject(Record):
|
||||
metadata = Metadata()
|
||||
schema_name = String() # Which schema this object belongs to
|
||||
values = Map(String()) # Field name -> value
|
||||
confidence = Double()
|
||||
source_span = String() # Text span where object was found
|
||||
```
|
||||
|
||||
### 4. Enhanced Knowledge Schemas
|
||||
|
||||
#### 4.1 Object Embeddings Enhancement
|
||||
Update `knowledge/embeddings.py` to support structured object embeddings better:
|
||||
|
||||
```python
|
||||
class StructuredObjectEmbedding(Record):
|
||||
metadata = Metadata()
|
||||
vectors = Array(Array(Double()))
|
||||
schema_name = String()
|
||||
object_id = String() # Primary key value
|
||||
field_embeddings = Map(Array(Double())) # Per-field embeddings
|
||||
```
|
||||
|
||||
## Integration Points
|
||||
|
||||
### Flow Integration
|
||||
|
||||
The schemas will be used by new flow modules:
|
||||
- `trustgraph-flow/trustgraph/decoding/structured` - Uses StructuredDataSubmission
|
||||
- `trustgraph-flow/trustgraph/query/nlp_query/cassandra` - Uses NLP query schemas
|
||||
- `trustgraph-flow/trustgraph/query/objects/cassandra` - Uses structured query schemas
|
||||
- `trustgraph-flow/trustgraph/extract/object/row/` - Consumes Chunk, produces ExtractedObject
|
||||
- `trustgraph-flow/trustgraph/storage/objects/cassandra` - Uses Rows schema
|
||||
- `trustgraph-flow/trustgraph/embeddings/object_embeddings/qdrant` - Uses object embedding schemas
|
||||
|
||||
## Implementation Notes
|
||||
|
||||
1. **Schema Versioning**: Consider adding a `version` field to RowSchema for future migration support
|
||||
2. **Type System**: The `Field.type` should support all Cassandra native types
|
||||
3. **Batch Operations**: Most services should support both single and batch operations
|
||||
4. **Error Handling**: Consistent error reporting across all new services
|
||||
5. **Backwards Compatibility**: Existing schemas remain unchanged except for minor Field enhancements
|
||||
|
||||
## Next Steps
|
||||
|
||||
1. Implement schema files in the new structure
|
||||
2. Update existing services to recognize new schema types
|
||||
3. Implement flow modules that use these schemas
|
||||
4. Add gateway/rev-gateway endpoints for new services
|
||||
5. Create unit tests for schema validation
|
||||
Loading…
Add table
Add a link
Reference in a new issue