Aweh

F-1 immigration guidance is fragmented across dense primary sources, while generic chat answers can sound confident without being grounded or appropriately limited.

I wanted to explore whether retrieval, citations, and explicit escalation could make a domain-specific assistant more useful without pretending it replaces legal counsel.

• Official immigration PDFs (USCIS policy manual, 8 CFR, SEVP guidance) are chunked and embedded into Supabase pgvector for semantic retrieval.
• Hybrid retrieval merges dense vector search with keyword matching, then Cohere rerank selects the most relevant spans before the agent composes an answer.
• User profile context rewrites queries for personalization; the agent is instructed to cite sources and flag when a question needs human legal counsel.

• Official documents are chunked and embedded into Supabase with pgvector.
• Dense and keyword retrieval are combined, then reranked before context reaches the model.
• Profile-aware query rewriting personalizes retrieval while answer rules require citations and escalation for ambiguous cases.

• Representative student questions test retrieval relevance across OPT, CPT, travel, and employment scenarios.
• Grounded answers are reviewed for citation support and appropriate uncertainty.

• Retrieval quality, not response fluency, is the first bottleneck in high-stakes assistants.
• A useful system needs a visible boundary between information and individualized advice.

• How should conflicting or superseded source documents be represented?
• What evaluation rubric best captures both usefulness and responsible escalation?