Public corpus
Safe Markdown documents are collected, cleaned, section-chunked, and exported with traceable metadata.
Transformer Project 10 · Vercel Capstone
AI Portfolio RAG Assistant
Ask evidence-backed questions about machine learning, NLP, computer vision, deployment, and portfolio experience. Every answer exposes retrieved evidence, source citations, relevance scores, and runtime latency.
complete public corpus220 documents3157 chunks
sentence-transformers/all-MiniLM-L6-v2
Real Transformer embeddings are active.Retriever + grounded generator
Search the portfolio
Use filters or ask a natural-language question. Answers are limited to indexed public evidence.
51/700 · Ctrl/Cmd + Enter
Transformer retrieval
Precomputed MiniLM or E5 embeddings are combined with lexical matching for efficient Vercel retrieval.
Grounded generation
A hosted instruction model can generate cited answers, while the extractive composer remains a safe fallback.
Evidence and evaluation
The interface exposes source chunks and latency, while offline scripts measure retrieval and answer quality.
Measured, not claimed
Portfolio evaluation
These values are generated from committed evaluation artifacts after the corpus and models are run locally.
Evaluation statusmeasured
Best retrieverminilm-plus-cross-encoder
Recall@50.646
nDCG@50.596
Groundedness0.025
Citation precision0.000
Refusal accuracy0.875
Quality gates4/9
Local P95 latency348.3 ms
Retrieval qualityHit Rate@K, Precision@K, Recall@K, MRR, MAP, and nDCG@K.
Answer groundednessClaim-level NLI support against retrieved evidence.
Citation correctnessCitation precision, completeness, and unsupported-claim rate.
Response latencyEmbedding, retrieval, generation, total, median, P90, and P95.
Pending values remain visibly labeled until the full evaluation notebook has been run. No placeholder result is presented as a final score.