Docs / Get started

Requirements and settings

Most problems with answers about your systems come from three settings left at values meant for short writing tasks. Set them once, check the model, and Verbiqo is ready.

At a glance

All three are in Settings, under the model. In a terminal they are VERBIQO_CONTEXT_TOKENS, VERBIQO_TIMEOUT_MS and VERBIQO_TEMPERATURE.

Context window

How much text the model can read at once: your question, the conversation so far, and what Verbiqo measured. About a third of it is given to evidence, up to 24,000 tokens, so a window above about 68,000 tokens adds nothing more for investigations.

Leave the setting blank and Verbiqo asks the model server:

When nothing reports it, Verbiqo assumes 4,096 tokens, the smallest common size. Answers about your systems then rest on a fraction of what was measured. The sign: Evidence was reduced to fit a …-token budget at the end of an answer, with a small number. Set Context window to what the server really has.

Timeout

How long Verbiqo waits for the model, from 1 to 300 seconds, entered in milliseconds. For a Prometheus question it is also the limit for the whole investigation: every query and every model step together.

• A question about a database, Redis, Elasticsearch or Kubernetes starts with a planning step: one short model call that says what the question needs, with its own 20-second limit. If the model does not answer in time, Verbiqo plans by its own rules and carries on. The reads and the answer then follow, each within the timeout.
• A root-cause question over a day runs 20 to 40 queries and several model steps. At the 30-second default it stops early and says so; set 300000.
• A reasoning model, or a busy shared server, needs the most.
• Behind a gateway, the gateway’s own timeout must be at least as long, or it ends the request first (LiteLLM: timeout, below).

Temperature

How much the model’s wording and choices vary from one run to the next. Verbiqo’s figures come from what it measured either way; temperature changes which metrics the model chooses and how it explains them.

• 0.2 keeps choices steady and still reads naturally.
• Blank sends none, and the server’s default applies. Ollama’s is 0.2 in Verbiqo; vLLM uses the model’s own, 0.7 for Qwen3, and the same question can then be answered differently twice.
• The range is 0 to 2 (Anthropic: 0 to 1). A few newer hosted models refuse any temperature; leave it blank for those.

It can be set on the server instead, so every computer gets it: vLLM --override-generation-config '{"temperature": 0.2}', llama.cpp --temp 0.2.

The model server

• Context: 32,768 tokens per request (vLLM --max-model-len 32768; llama.cpp -c 32768 per slot).
• Tool calling on: vLLM --enable-auto-tool-choice --tool-call-parser hermes for Qwen; llama.cpp --jinja. Without it Verbiqo still works, more slowly.
• No long reasoning: an instruct model, or a hybrid one with reasoning off. See Reasoning models.
• A key with --api-key, and the same key in Settings.

Two set-ups tested and working, with their exact commands: Tested and working.

Behind LiteLLM

model_list:
  - model_name: qwen3-30b-instruct
    litellm_params:
      model: hosted_vllm/qwen3-30b-instruct
      api_base: http://gpu-server:8000/v1
      api_key: os.environ/VLLM_API_KEY
      timeout: 600
      temperature: 0.2
    model_info:
      max_input_tokens: 32768

• model_info.max_input_tokens is what lets Verbiqo learn the context window through the gateway.
• timeout is in seconds; keep it above Verbiqo’s.
• The model_name is what you enter as Model in Verbiqo. The model check warns when the gateway answers with a different model.

Your systems

• A read-only account for each system: PostgreSQL, MySQL, SQL Server, Redis, Elasticsearch and OpenSearch, Kubernetes, Prometheus, Splunk.
• Reachable from your computer: Verbiqo connects from where it runs, so the database port, the Kubernetes API server, Prometheus or Splunk must be reachable from there (VPN, firewall).
• Your organization’s certificates: name the CA once in Settings › Trusted CA (below Endpoint); it covers the AI service too. A terminal on the same computer trusts it too, along with everything in the computer’s trust store.
• Prometheus: press Build Index once after connecting, so Verbiqo knows its jobs, hosts and metrics. A full build needs an embedding service: an OpenAI-compatible embeddings endpoint on this computer (such as Ollama with nomic-embed-text) or an approved HTTPS one, returning at most 2,048 dimensions. Embedding service requirements.

Prometheus needs its local map before the first question: a full build (up to half an hour on a large server, with an embedding service) or a refresh per technology (seconds, no service needed). It decides how well questions are scoped; without it, each question discovers live and the answers are slower and weaker. The local map.

Check it works

1. The model answers

   In Settings, Test & Load Models, then choose the model.

2. The model is fit

   Run the model check once, and keep the report.

3. Each system connects

   Extensions…, then Test Connection for each.

4. A first question

   Attach a system and ask something simple: is anything down right now? If the answer ends with Evidence was reduced… or stopped early, revisit the context window and timeout above.