askUOS/docs/backend_config_example.yaml at main · virtUOS/askUOS · 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
models:

  - provider: "google" # ["openai", "google", "self-hosted"]
    role: "main" # ["main", "helper", "subagent"] # main: model used for agentic tasks and generatiom. Helper: used for summary generation and tasks that do not require reasoning (easy tasks). Subagent: used for ephemeral MCP subagent tool-calling tasks (see mcp_agents below)
    model_name: "gemini-3-flash-preview"
    # timeout: 60 # Optional. LLM client request timeout in seconds (default: 60)

  - provider: "google" # ["openai", "google", "self-hosted"]
    role: "helper" # ["main", "helper", "subagent"]
    model_name: "gemini-2.5-flash-lite"

  # Optional. If omitted, MCP subagents (mcp_agents below) reuse the "main"
  # model above. Configure this only if you want subagent tool-calling tasks
  # to use a different (e.g. cheaper/faster) model than the main conversation.
  # - provider: "openai" # ["openai", "google", "self-hosted"]
  #   role: "subagent" # ["main", "helper", "subagent"]
  #   model_name: "gpt-4.1-mini"

language: "Deutsch"  # Literal["Deutch", "English"].

redis:
  host: "redis"
  port: "6379"
  username: ""
  password: ""
  ttl_graph_cache: 120 # how long messages are cached. Msgs older than this value won't be shown to the user
  connection_timeout: 15 # Optional. Seconds to wait for a Redis connection from the pool (default: 15)

# General system messages are in the locale. (Optional)
messages:
  - msg_name: "further_help" # Literal["further_help"]
    english: |
      If you need personal assistance regarding studying at the university, you can visit this wesbsite ...  or ... .
    german: |
      Wenn Sie persönliche Unterstützung zum Studium an der Universität benötigen, können Sie die Website des ...  oder ... .

# the faq collection needs a metadata field. Name it "url", this link will be shwon to the user so that they can double check the information provided by the chatbot.
graph:
  summary_threshold: 70000 # in tokens
  faq:
    activate: True
    collection_name: "faq"
  # This tool is a "standalone tool" (not MCP-based), so its description
  # lives here in backend_config.yaml rather than in prompts/ -- it's only
  # needed at all if activate: True. See docs/getting-started.md
  skip_grading_for_high_trust_sources: True
  # Max input context (in tokens) of the embedding model below -- used to
  # size how much of the retrieved documents get embedded for
  # grade_documents' similarity pre-filter. Lower this if a smaller
  # embedding model is configured.
  embedding_prefilter_max_context_tokens: 8192
  # Optional. Embedding-similarity pre-filter thresholds for grade_documents
  # (applies to the remaining, non-high-trust path). Unset disables the
  # pre-filter; requires the top-level `embedding:` block to be configured.
  embedding_prefilter_high_threshold: 0.6
  embedding_prefilter_low_threshold: 0.01
  troubleshooting:
    collection_name: "troubleshooting"
    activate: False # set True to enable the troubleshooting tool; description below is then required
    description: |
      Describe here what this tool is for and when the agent should use it
      (e.g. which software/product it can answer technical questions about).

application:
  debug: False
  tracing: False
  recursion_limit: 8 # default is 12 https://langchain-ai.github.io/langgraph/how-tos/recursion-limit/
  # Optional. Origins allowed to call this API directly from browser-based
  # JS (CORS) -- e.g. a standalone React/JS frontend, or a chat widget
  # embedded on another domain. Leave empty (default) if the only caller is
  # the bundled Streamlit UI or a server-to-server integration; CORS is a
  # browser-only mechanism and doesn't apply to those.
  # cors_allowed_origins:
  #   - "https://your-frontend.example.edu"

# Only needed if using Milvus vector db (or grade_documents' embedding
# pre-filter). Self-hosted, OpenAI-compatible endpoint -- API key must be
# provided via API_KEY_SELF_HOSTED_EMBEDDING in the env file.
embedding:
  model_name: "BAAI/bge-m3"
  base_url: "https://my-litellm-instance.de"
  timeout: 3

vector_db_settings:
  # Vector database configuration
  type: "Infinity-RAGFlow"
  settings:
    base_url: "https://ragflow.de"
    chunk_size: 10
    # Optional RAGFlow client timeouts, in seconds (defaults shown below).
    # connect_timeout: 15.0 # Time to establish connection
    # read_timeout: 60.0    # Time to receive response (RAG can be slow)
    # write_timeout: 15.0   # Time to send request body
    # pool_timeout: 5.0     # Time waiting for connection from pool


# Required. Web crawler configuration (crawl4ai), used by the web search tool.
crawl_settings:
  base_url: "http://crawl4ai:11235/crawl" # crawl4ai service, as wired in docker-compose.yml
  ttl_redis: 172800 # 48h default TTL (how long a page is cached before it's crawled again)
  max_links: 6 # Optional. Max number of web pages visited per search (default: 6)
  crawl_payload: # doc: https://www.postman.com/pixelao/pixel-public-workspace/documentation/c26yn3l/crawl4ai-api
    browser_config:
      type: "BrowserConfig"
      params:
        headless: true
    crawler_config:
      type: "CrawlerRunConfig"
      params:
        stream: false
        cache_mode:
          type: "CacheMode"
          params: "enabled"
        check_cache_freshness: true
        cache_validation_timeout: 10.0
        word_count_threshold: 100
        target_elements:
          - "main"
          - "div#content"
        scan_full_page: true

# Optional. If set up, implement e.g., cron job to run /app/src/chatbot_log/log_cleanup.py
log_settings:
  delete_logs_days: 90 # Number of days to keep logs before deletion