Pakistan North Travel AI
Ask about northern Pakistan and see the pages behind each answer.

Ask the travel guide
Where can I refuel before Khunjerab?
Refuel at Sost.
Travel guide · PDF pages 6–8
Tools and Skills
Technologies
- Streamlit
- LangChain
- Gemini
- LangGraph
- Chroma
Skills Applied
- Conversational RAG
- Page-level citations
- Passage indexing
- Retrieval evaluation
The problem
The useful details in a travel guide are rarely all on one page. A question about a route can depend on transport, fuel, and terrain. Travellers also need a way to check where an answer came from, especially when they’re planning around a constraint.
What I built
I turned a 92-page guide into 49 searchable passages, preserving page ranges and extracting fare tables alongside the text. The Streamlit app uses Gemini, LangChain, LangGraph, and a bundled Chroma index. It supports follow-up questions, shows the original passages, and offers a trip planner and beta destination comparisons.
Why I built it this way
Compare retrieval methods before adding complexity.
I evaluated dense similarity, BM25 keyword search, and hybrid retrieval. Dense remained the default and retrieved complete evidence in the top five for 30 of 31 eligible held-out cases.
Carry context forward and keep citations inspectable.
LangGraph keeps recent turns in the session so a follow-up can refer to the last destination or preference. Source IDs resolve to stored passages and page ranges users can open.
Keep recommendation searches bounded.
The beta can make up to two targeted searches for a comparison. Its answers distinguish what the guide documents, what suitability the model infers, and what still needs checking.
What I measured
On 32 answerable held-out questions, dense retrieval achieved 93.75% Hit@1 and 100% Hit@5. Those figures measure whether the system found relevant evidence.
In a separate development evaluation, three of six recommendation cases fully passed. One case cited useful pages but recommended a 3–4-hour walk to someone who asked for minimal walking. That result shows why I evaluate the recommendation itself as well as the citation. The app draws on the travel guide and provides no live road, weather, or booking data.
Have a knowledge base people struggle to search?
Bring a few questions and the documents that answer them. We can use those examples to define how a first version should retrieve, cite, and handle gaps.
Book a Discovery Call