Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/tudoumono/Sherpa/llms.txt

Use this file to discover all available pages before exploring further.

Sherpa lets you narrow the documents it searches and choose the AI model that reasons over them. These two controls — scope paths and the AI (Brain) selector — are independent and can be adjusted at any point in a conversation. Scope paths limit search results to a specific subfolder; the AI selector determines which provider generates the answer. Both settings are scoped to the current user and persist between sessions.

​
Scope paths

When knowledge retrieval is turned on, Sherpa searches the entire registered document world by default. You can narrow that search to a specific subfolder using the scope panel. Scope works as a folder prefix filter. Selecting 02_設計 limits all impact analysis, troubleshooting searches, and spec lookups to documents whose path starts with that prefix. You can select multiple folders at once to span several subtrees simultaneously.
Common shared components — standard copybooks, cross-cutting standards documents, and similar shared resources — are included in every search regardless of which scope paths you have selected. You will not accidentally miss a shared dependency by narrowing scope.
After each answer, the scope actually used is shown in the answer envelope under 参照中の範囲 (“scope in use”). This reflects either your explicit selection or, if you left scope unset, the inferred coverage.
  1. Make sure the ナレッジ参照 (Knowledge retrieval) toggle is on — the scope selector is hidden when knowledge retrieval is off.
  2. Click the 範囲 (Scope) button near the input bar. A folder tree panel opens.
  3. Click 📂 全体（この取込ディレクトリすべて） to clear any selection and search the entire world.
  4. Click any folder row to select it. Selected folders are highlighted. Click again to deselect.
  5. You can select multiple folders. The scope chip above the input bar updates to show your selection.
  6. Send your question. The answer envelope will confirm which scope paths were applied.
If your registered world contains only one leaf-level folder (or only one world is registered), the scope selector is hidden automatically — there is nothing meaningful to filter.

​
AI model selector

The 頭脳 (Brain) badge near the message input shows the currently active AI provider. Click it to open the provider menu. Your selection is saved via PUT /settings and applied to all subsequent queries in the session.
ProviderDefault modelKey / Config env varNotes
Codexgpt-5.5OPENAI_API_KEYAgentic mode: Codex runs its own grep and file-read tools inside a sandboxed environment. The thinking trace in the right pane shows each step. Best for complex dependency traversals and authoring tasks. This is the recommended default.
OpenAI APIgpt-5.5 (configurable)OPENAI_API_KEYStandard chat completion. Only document text is sent to the API — original files are never uploaded externally. Model ID is configurable per user.
Geminigemini-2.5-flash (configurable)GEMINI_API_KEYGoogle’s Gemini API. Model ID is configurable per user. Same text-only data boundary as OpenAI.
Ollama (Local LLM)qwen2.5 (configurable)Base URL configurableRuns inference against a local Ollama server. No data leaves your network. Configure the base URL and model ID in Settings.
Heuristic (No AI)—None requiredTemplate-based fallback. No external API calls. Fastest response, but answers are less nuanced. Used when no AI provider is configured.

​
Codex reasoning strength

When Codex is the active provider, you can set its reasoning strength in Settings:
LevelBehaviour
low (default)Fastest and cheapest. Suitable for most queries.
mediumMore thorough traversal; better for queries that span many files.
highDeepest analysis; highest cost and latency. Use for complex impact analyses across large codebases.

​
Configuring provider settings

Open ⚙ APIキー等の詳細設定 from the Brain menu, or navigate to the Settings page directly. Changes are saved immediately via PUT /settings and take effect on the next query. 
PUT /settings
Content-Type: application/json

{
  "agent": "codex",
  "codex_reasoning": "medium",
  "codex_model": "gpt-5.5",
  "openai_api_key": "sk-...",
  "openai_model": "gpt-5.5",
  "gemini_api_key": "AIza...",
  "gemini_model": "gemini-2.5-flash",
  "ollama_url": "http://localhost:11434",
  "ollama_model": "qwen2.5"
}
All fields are optional — omit any field to keep the current value. API keys are stored encrypted and are never returned in GET /settings responses (a boolean openai_key_set / gemini_key_set flag is returned instead). You can test a provider configuration before saving it: 
POST /settings/test
Content-Type: application/json

{
  "provider": "gemini",
  "gemini_api_key": "AIza...",
  "gemini_model": "gemini-2.5-flash"
}
The test endpoint accepts openai, gemini, ollama, or codex as the provider value.

Build docs developers (and LLMs) love

Get started for freeTalk to us