SLURP: Add comprehensive LightRAG query parameter support #14

New Issue

tony · 2025-10-04T02:06:48Z

tony commented

2025-10-04 02:06:48 +00:00

Feature Request: Enhanced RAG Query Parameter Support

Current State

SLURP currently only configures basic RAG integration parameters:

RAGEndpoint: URL to RAG service
RAGTimeout: Request timeout duration

See IntelligenceConfig in pkg/slurp/slurp.go:209-231

Problem

SLURP cannot leverage the full capabilities of LightRAG's query API. It uses only default parameters, which limits:

Query result quality control
Performance optimization
Context generation flexibility
Resource usage tuning

Requested Enhancement

Add comprehensive support for LightRAG query parameters:

1. Response Format Control

response_type: Control output format (e.g., "Multiple Paragraphs", "Single Paragraph", "Bullet Points")
Allows SLURP to request structured vs. narrative responses based on context needs

2. Knowledge Graph Retrieval

top_k: Number of KG entities/relationships to retrieve (default: 60)
Fine-tune graph traversal depth for different query types

3. Chunk Retrieval

chunk_top_k: Number of text chunks to retrieve (default: 10)
Balance between context breadth and token efficiency

4. Token Budget Management

max_entity_tokens: Maximum tokens for entity context
max_relation_tokens: Maximum tokens for relationship context
max_token_for_text_unit: Maximum tokens per text fragment
max_token_for_local_context: Local context token limit
max_token_for_global_context: Global context token limit
Critical for managing LLM context windows and costs

5. Reranking Control

enable_rerank: Toggle reranking for result quality
min_rerank_score: Minimum score threshold for inclusion

6. Query Modes

mode: Search strategy ("mix", "local", "global", "hybrid", "naive", "semantic", "keyword")
Different contexts may benefit from different retrieval strategies

7. Output Control

only_need_context: Return raw context without LLM generation
only_need_prompt: Return generated prompt without execution
Useful for debugging and optimization

8. Conversation History

history_turns: Number of conversation turns to include
Support for multi-turn context resolution

Proposed Implementation

Option 1: Extend IntelligenceConfig

type IntelligenceConfig struct {
    // ... existing fields ...
    
    // RAG query parameters
    RAGResponseFormat      string  `yaml:"rag_response_format" json:"rag_response_format"`
    RAGTopK               int     `yaml:"rag_top_k" json:"rag_top_k"`
    RAGChunkTopK          int     `yaml:"rag_chunk_top_k" json:"rag_chunk_top_k"`
    RAGMaxEntityTokens    int     `yaml:"rag_max_entity_tokens" json:"rag_max_entity_tokens"`
    RAGMaxRelationTokens  int     `yaml:"rag_max_relation_tokens" json:"rag_max_relation_tokens"`
    RAGMaxTextUnitTokens  int     `yaml:"rag_max_text_unit_tokens" json:"rag_max_text_unit_tokens"`
    RAGMaxLocalContext    int     `yaml:"rag_max_local_context" json:"rag_max_local_context"`
    RAGMaxGlobalContext   int     `yaml:"rag_max_global_context" json:"rag_max_global_context"`
    RAGEnableRerank       bool    `yaml:"rag_enable_rerank" json:"rag_enable_rerank"`
    RAGMinRerankScore     float64 `yaml:"rag_min_rerank_score" json:"rag_min_rerank_score"`
    RAGMode               string  `yaml:"rag_mode" json:"rag_mode"`
    RAGHistoryTurns       int     `yaml:"rag_history_turns" json:"rag_history_turns"`
}

Option 2: Per-Query Override
Add optional parameters to GenerateContext:

type RAGQueryOptions struct {
    ResponseFormat     string
    TopK              int
    ChunkTopK         int
    MaxEntityTokens   int
    MaxRelationTokens int
    MaxTextUnitTokens int
    EnableRerank      bool
    MinRerankScore    float64
    Mode              string
    OnlyContext       bool
    OnlyPrompt        bool
}

func (s *SLURP) GenerateContext(ctx context.Context, path string, 
    options *GenerationOptions, ragOpts *RAGQueryOptions) (*ContextNode, error)

Benefits

Performance: Tune token limits to balance quality vs. cost
Quality: Use reranking for critical context generation
Flexibility: Different query modes for different use cases
Debugging: Access raw context and prompts for optimization
Cost Control: Precise token budget management
Adaptability: Match retrieval strategy to query type

Acceptance Criteria

SLURP config supports all LightRAG query parameters
Parameters can be set globally in config
Parameters can be overridden per-query
Sensible defaults match LightRAG recommendations
Documentation updated with parameter descriptions
Examples showing different query configurations
Tests verify parameter passing to RAG endpoint

pkg/slurp/slurp.go - Main configuration
pkg/slurp/intelligence/ - Context generation (when implemented)
LightRAG API integration layer

Priority

Medium-High - This enables SLURP to fully leverage RAG capabilities and provides critical control over context generation quality and cost.

Labels

bzzz-task - Task coordination
enhancement - New feature
slurp - SLURP subsystem
rag - RAG integration

## Feature Request: Enhanced RAG Query Parameter Support ### Current State SLURP currently only configures basic RAG integration parameters: - `RAGEndpoint`: URL to RAG service - `RAGTimeout`: Request timeout duration See `IntelligenceConfig` in `pkg/slurp/slurp.go:209-231` ### Problem SLURP cannot leverage the full capabilities of LightRAG's query API. It uses only default parameters, which limits: - Query result quality control - Performance optimization - Context generation flexibility - Resource usage tuning ### Requested Enhancement Add comprehensive support for LightRAG query parameters: #### 1. **Response Format Control** - `response_type`: Control output format (e.g., "Multiple Paragraphs", "Single Paragraph", "Bullet Points") - Allows SLURP to request structured vs. narrative responses based on context needs #### 2. **Knowledge Graph Retrieval** - `top_k`: Number of KG entities/relationships to retrieve (default: 60) - Fine-tune graph traversal depth for different query types #### 3. **Chunk Retrieval** - `chunk_top_k`: Number of text chunks to retrieve (default: 10) - Balance between context breadth and token efficiency #### 4. **Token Budget Management** - `max_entity_tokens`: Maximum tokens for entity context - `max_relation_tokens`: Maximum tokens for relationship context - `max_token_for_text_unit`: Maximum tokens per text fragment - `max_token_for_local_context`: Local context token limit - `max_token_for_global_context`: Global context token limit - Critical for managing LLM context windows and costs #### 5. **Reranking Control** - `enable_rerank`: Toggle reranking for result quality - `min_rerank_score`: Minimum score threshold for inclusion #### 6. **Query Modes** - `mode`: Search strategy ("mix", "local", "global", "hybrid", "naive", "semantic", "keyword") - Different contexts may benefit from different retrieval strategies #### 7. **Output Control** - `only_need_context`: Return raw context without LLM generation - `only_need_prompt`: Return generated prompt without execution - Useful for debugging and optimization #### 8. **Conversation History** - `history_turns`: Number of conversation turns to include - Support for multi-turn context resolution ### Proposed Implementation **Option 1: Extend IntelligenceConfig** ```go type IntelligenceConfig struct { // ... existing fields ... // RAG query parameters RAGResponseFormat string `yaml:"rag_response_format" json:"rag_response_format"` RAGTopK int `yaml:"rag_top_k" json:"rag_top_k"` RAGChunkTopK int `yaml:"rag_chunk_top_k" json:"rag_chunk_top_k"` RAGMaxEntityTokens int `yaml:"rag_max_entity_tokens" json:"rag_max_entity_tokens"` RAGMaxRelationTokens int `yaml:"rag_max_relation_tokens" json:"rag_max_relation_tokens"` RAGMaxTextUnitTokens int `yaml:"rag_max_text_unit_tokens" json:"rag_max_text_unit_tokens"` RAGMaxLocalContext int `yaml:"rag_max_local_context" json:"rag_max_local_context"` RAGMaxGlobalContext int `yaml:"rag_max_global_context" json:"rag_max_global_context"` RAGEnableRerank bool `yaml:"rag_enable_rerank" json:"rag_enable_rerank"` RAGMinRerankScore float64 `yaml:"rag_min_rerank_score" json:"rag_min_rerank_score"` RAGMode string `yaml:"rag_mode" json:"rag_mode"` RAGHistoryTurns int `yaml:"rag_history_turns" json:"rag_history_turns"` } ``` **Option 2: Per-Query Override** Add optional parameters to `GenerateContext`: ```go type RAGQueryOptions struct { ResponseFormat string TopK int ChunkTopK int MaxEntityTokens int MaxRelationTokens int MaxTextUnitTokens int EnableRerank bool MinRerankScore float64 Mode string OnlyContext bool OnlyPrompt bool } func (s *SLURP) GenerateContext(ctx context.Context, path string, options *GenerationOptions, ragOpts *RAGQueryOptions) (*ContextNode, error) ``` ### Benefits 1. **Performance**: Tune token limits to balance quality vs. cost 2. **Quality**: Use reranking for critical context generation 3. **Flexibility**: Different query modes for different use cases 4. **Debugging**: Access raw context and prompts for optimization 5. **Cost Control**: Precise token budget management 6. **Adaptability**: Match retrieval strategy to query type ### Acceptance Criteria - [ ] SLURP config supports all LightRAG query parameters - [ ] Parameters can be set globally in config - [ ] Parameters can be overridden per-query - [ ] Sensible defaults match LightRAG recommendations - [ ] Documentation updated with parameter descriptions - [ ] Examples showing different query configurations - [ ] Tests verify parameter passing to RAG endpoint ### Related Components - `pkg/slurp/slurp.go` - Main configuration - `pkg/slurp/intelligence/` - Context generation (when implemented) - LightRAG API integration layer ### Priority **Medium-High** - This enables SLURP to fully leverage RAG capabilities and provides critical control over context generation quality and cost. ### Labels - `bzzz-task` - Task coordination - `enhancement` - New feature - `slurp` - SLURP subsystem - `rag` - RAG integration

tony added the

bzzz-task

enhancement

labels 2025-10-04 02:07:35 +00:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: tony/CHORUS#14