RAGushka/config.example.yaml at main · baroque-cat/RAGushka · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
logging:
  # Enable debug mode without log interceptor
  debug_mode: true
  # List of loggers to suppress in normal mode
  suppress_loggers:
    - "torch"
    - "torch.dynamo"
    - "torch._dynamo"

core:
  # Global HTTP timeout for connection pool (seconds)
  global_http_timeout: 600.0
  # Timeout for health check requests (seconds)
  health_check_timeout: 5.0
  # Retry Configuration for API calls
  max_retries: 3
  retry_min_wait: 1.0
  retry_max_wait: 10.0

multimodal_pipeline:
  # Parsing method
  parse_method: auto
  # Extract images (VLM)
  enable_image_processing: true
  # Extract tables (VLM/LLM)
  enable_table_processing: true
  # Extract equations
  enable_equation_processing: true
  # Context window around multimedia
  context_window: 2
  # Context unit (page or chunk)
  context_mode: page
  # Maximum tokens in context
  max_context_tokens: 3000
  # Include headers in context
  include_headers: true
  # Include captions for images/tables
  include_captions: true

markdown_pipeline:
  # Maximum tokens per chunk for docling semantic chunking
  max_tokens: 4000

server:
  # Server host address
  host: "localhost"
  # CORS allowed origins
  cors_origins: "*"
  # JWT signing algorithm
  jwt_algorithm: "HS256"
  # JWT token expiration time in hours
  token_expire_hours: 24
  # Guest token expiration time in hours
  guest_token_expire_hours: 1
  # Enable automatic token renewal
  token_auto_renew: true
  # Renew token when remaining time < this ratio of total
  token_renew_threshold: 0.3
  # Comma-separated paths that don't require authentication
  whitelist_paths: "/health,/docs,/redoc,/openapi.json,/,/chat,/knowledge-graph,/api,/assets,/auth-status"

minio:
  # Maximum file size in bytes to download for Base64 encoding (default 20MB)
  max_file_size_bytes: 20971520

llm:
  # LLM endpoint (without /chat/completions)
  base_url: https://url/v1
  # Text model name
  model_name: llm
  # Enable thinking mode for Qwen models
  enable_qwen_thinking: false
  # Timeout for LLM API calls (seconds)
  timeout: 360.0

vlm:
  # VLM endpoint
  base_url: https://url/v1
  # Visual model name
  model_name: vlm
  # Enable thinking mode for Qwen-VL models
  enable_qwen_thinking: false
  # Timeout for VLM API calls (seconds)
  timeout: 360.0

embedding:
  # Embedding endpoint
  base_url: https://url/v1
  # Embedding model name
  model_name: emb
  # Maximum tokens per embedding request
  max_tokens: 8192
  # Vector dimension (provider-dependent)
  dimension: 2048
  # Maximum batch size for embeddings (optional, provider-dependent; omit or set to None for universal support)
  batch_num: 10
  # Enable asymmetric embeddings (text_type='query' for search, text_type='document' for indexing).
  asymmetric: false
  # Timeout for Embedding API calls (seconds)
  timeout: 60.0

rerank:
  # Enable/disable reranking service
  enabled: false
  # Reranker API base URL
  base_url: "https://url/v1/reranks"
  # Reranker model name
  model_name: "rerank-model"
  # Maximum documents per request (API limit)
  max_docs_per_request: 500
  # Maximum tokens per document (API limit)
  max_tokens_per_doc: 4000
  # Maximum tokens per request (API limit)
  max_tokens_per_request: 120000
  # Whether the model supports dynamic instructions (asymmetric reranking)
  supports_instruct: true
  # Minimum score threshold (0.0 - 1.0) for keeping chunks after reranking
  min_score: 0.0

retrieval:
  # Maximum number of entities/relations retrieved from vector DB
  top_k: 60
  # Maximum number of text chunks retrieved from vector search (before reranking)
  vector_search_top_k: 150
  # Maximum number of text chunks kept after reranking
  chunk_top_k: 20
  # Minimum cosine similarity threshold for vector search (filters out noise)
  cosine_threshold: 0.2
  # Maximum token budget for the entire prompt context
  max_total_tokens: 30000
  # Maximum token budget allocated for entities
  max_entity_tokens: 6000
  # Maximum token budget allocated for relations
  max_relation_tokens: 8000

search_strategies:
  # Default QA search strategy
  qa:
    # Enable/disable QA search strategy
    enabled: true

  # Similarity search strategy
  similarity:
    # Enable/disable Similarity search strategy
    enabled: false