v2.1.0 - Bug fixes and observability improvements

Fixes

• Fix WikiVoyage API 403 by adding required User-Agent header (#12)
• Fix raw markdown rendering in PDF activities section (#13)
• Remove json_mode warning when using Claude as LLM provider (#14)

Improvements

• Add per-destination logging in enrich node for better pipeline observability (#14)
• Add RAG retrieval logging in suggest node (#14)
• Add test verifying User-Agent header is sent (#12)

Testing

• 196 tests, all passing

v2.0.0 - Multi-provider LLM, RAG, MCP Server

What's New

Multi-provider LLM Support

• Switch between Claude and OpenAI via LLM_PROVIDER env var
• Provider-agnostic factory pattern -- same code, different models

RAG-Enhanced Recommendations

• Qdrant vector store with embedded mode (no external server needed)
• WikiVoyage knowledge base: scrape, chunk, embed, and index 35+ destinations
• User interaction history for personalized suggestions across sessions
• Prompt injection sanitization for safe RAG context injection

MCP Server

• Expose the flight report pipeline as an MCP tool
• Compatible with Claude Code, Claude Desktop, and any MCP client

Claude Code Skill

• Natural language triggers for travel queries
• Seamless integration with the MCP server

Pipeline Changes

• New store_interaction node saves completed requests for future personalization
• suggest and enrich nodes query Qdrant for context before LLM calls
• Graceful fallback when RAG is unavailable

Testing

• 196 tests (103 new), all passing
• End-to-end integration tests with real Qdrant
• All external APIs mocked

Setup

uv sync
uv run ingest-wikivoyage    # One-time knowledge base ingestion
uv run uvicorn src.app.main:app --reload

v1.0.0 - LangGraph Travel Recommendation Agent

Initial Release

LangGraph-based travel recommendation agent that generates comparative PDF reports with flight options.

Features

• Natural language intake via GPT-4o with structured output
• 5-stage pipeline: intake -> validate -> suggest -> search_flights -> enrich -> generate_report
• Real flight search via Amadeus API with budget filtering
• Weather enrichment via Open-Meteo (previous year data as proxy)
• LLM-generated activity recommendations per destination
• Weighted scoring algorithm (climate 30%, activity 30%, price 25%, stops 15%)
• PDF report generation with comparison table and detail pages
• SSE streaming endpoint for real-time pipeline progress
• PDF download route
• All prompts and user-facing content in Spanish

API Endpoints

• POST /api/chat -- Generate travel report
• POST /api/chat/stream -- SSE streaming version
• GET /api/reports/{filename} -- Download generated PDF
• GET /api/graph/viewer -- Interactive Mermaid pipeline diagram

Tech Stack

Python 3.12+, LangGraph, FastAPI, GPT-4o, Amadeus API, ReportLab

Testing

93 tests covering all pipeline nodes, services, and API endpoints