4 · 2026
Tenerife travel assistant
A chat over a PDF travel guide. It retrieves with ChromaDB, answers with GPT-4.1-mini and tool calling while citing its source, and ships as a Streamlit app built on a reusable Python package.
- PDF guideTENERIFE.pdf, extracted with PyPDF
- Chunkingfixed length, 200-character overlap
- Embeddingstext-embedding-3-small
- ChromaDBpersisted to disk
Input
- Chat10-turn history, 1,000 characters max
Orchestration
- GPT-4.1-minitool calling, 5 rounds max
Tools
- search_tenerife_infotop-k passages from ChromaDB
- get_weathersimulated weather
Answer
- Answer with sourcethe model writes from the context and cites the chunk
Problem
Answer questions about tourism in Tenerife using only a PDF guide, so that the assistant cites where each fact comes from and says explicitly when the guide doesn’t cover something.
Approach
The project goes through three stages, each built on the previous one:
- Exploration notebook: PDF extraction, overlapping chunking,
text-embedding-3-smallembeddings and a ChromaDB collection persisted to disk. tenerife_ragpackage: the notebook logic refactored into modules (config,vector_store,tools,tool_loop) that can be imported and tested on their own.- Streamlit app: a multi-turn chat that reuses the package and the already-persisted database, with no re-ingestion of the PDF.
Tool calling instead of an agent framework. The assistant only needs two capabilities:
search_tenerife_info (guide search) and get_weather (simulated weather). A custom loop capped at 5 rounds, with
arguments validated against JSON Schema (strict: true) and a log of every execution, gives more control, lower
latency and simpler debugging than adding LangChain or AutoGen.
Also: temperature 0.2 to stay close to the context, history capped at 10 turns, input validation (1,000
characters max) and output sanitising that strips technical logs while keeping source attribution.
Results
- Answers grounded in the guide, each with its attribution (
Source: TENERIFE.pdf (chunk N, score S)), and an explicit refusal when the fact isn’t there. - Correct tool selection depending on the question.
- Coherent multi-turn conversation: it resolves references like “and what can I eat there?” after talking about Mount Teide.
- Argument errors caught without interrupting the session.
What I learned
Sometimes the best agent architecture is not having one. For two tools and a narrow domain, an explicit tool-calling loop is more predictable and cheaper. And while a notebook proves something works, only a package with clear boundaries makes it reusable.
A formal evaluation (RAGAS or an LLM judge) that puts numbers on faithfulness and relevance is still pending, and is documented as a limitation.