Pilgrim Companion
AI guidance for Hajj and Umrah. A local-first question-answering application that retrieves multilingual sources and grounds responses in cited context.
Retrieval pipeline
A simplified view of the pipeline described in the project README.
- English or Arabic question
- Domain routing
- Dense and BM25 retrieval
- Rank fusion and relevance checks
- Optional reranking
- Grounded prompt and local or API model
- Streamed response with sources
Retrieval evaluation
Results reported in the repository for a 12-question English and Arabic regression set on a CPU development machine. These results are not a measure of religious correctness or production accuracy.
| Metric | Result |
|---|---|
| Retrieval Recall@4 | 100% |
| Mean Reciprocal Rank | 0.926 |
| Out-of-domain rejection | 100% |
| Median retrieval latency | 51.8 ms |
| Knowledge-base chunks | 123 unique chunks |
- Combines semantic and keyword retrieval for English and Arabic questions.
- Supports local Ollama models and OpenAI-compatible APIs.
- Includes source citations, streaming responses, and out-of-domain handling.
- Uses conversation-safe caching for standalone questions to avoid reusing cached results for follow-up questions.