Sherpa is an agentic RAG platform that lets teams register internal document folders — design specs, COBOL/JCL source, Office files, Markdown — and query them through a chat interface backed by AI agents, Elasticsearch full-text search, and a Neo4j knowledge graph. Every answer includes citations that link directly to the original file for download.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.
Quickstart
Get Sherpa running locally in under five minutes and send your first chat query.
Core Concepts
Understand worlds, scopes, lenses, and the directory-mirror ingest model.
Chat Interface
Learn how to use impact analysis, troubleshooting, and spec Q&A from the chat UI.
API Reference
Explore the full FastAPI surface — endpoints, request shapes, and response schemas.
What Sherpa does
Sherpa provides three analysis lenses, all accessible from a single chat input:Impact Analysis
Ask “what does changing X affect?” and get a ranked list of impacted modules, design docs, and COBOL programs with traversal paths and downloadable evidence.
Troubleshooting
Describe a symptom and get confidence-ranked root-cause candidates, each with supporting citations from incident records and design specs.
Spec Q&A
Ask “how is billing calculated?” and get an excerpt from the relevant spec with a link to the original file. If there is no evidence, Sherpa says so explicitly.
How it works
Register a folder
An admin registers a local or network directory as a world. Sherpa mirrors the folder tree into Elasticsearch and Neo4j on each ingest run.
Ask in chat
Users open the chat UI at
http://127.0.0.1:8000/ui/chat.html, turn on Knowledge retrieval, and type their question. Sherpa selects the right search path automatically.Review cited answers
The answer arrives with source citations. Click any citation to download the original Office binary, COBOL source file, or Markdown document.
Key capabilities
- AI-agent backbone — Codex (gpt-5.5), OpenAI GPT-4o, Gemini, and local Ollama models; switchable per user
- Three search paths — raw grep (exact match), Elasticsearch BM25, and Neo4j graph traversal
- Scope filtering — narrow any query to a subfolder prefix; common components merge automatically
- Personal workspace — each user uploads private reference files that only they can grep; never indexed in the shared RAG
- Conversation sharing — share conversations with specific users, with expiry dates and sanitized-snapshot support for privacy
- Tamper-evident audit log — every admin action is logged with a SHA-256 hash chain for integrity verification
- Production-ready auth — opt-in login (
SHERPA_AUTH_ENABLED=1), role-based access (admin/user), and session cookies
Sherpa requires Python 3 and Docker (for PostgreSQL, Neo4j, and Elasticsearch). The application core runs directly on the host — not in a container — so that Codex sandbox isolation and WSL
/mnt access work correctly.