inputs/retrieval_pairs.json

[
  {
    "query": "Which city is the capital of France?",
    "document": "Paris is the capital and most populous city of France."
  },
  {
    "query": "What is Canada's national capital rather than its largest city?",
    "document": "Ottawa is the capital of Canada; Toronto is the country's largest city."
  },
  {
    "query": "How do green plants turn light into stored chemical energy?",
    "document": "Photosynthesis uses light energy to convert carbon dioxide and water into sugars and oxygen."
  },
  {
    "query": "Which organelle produces most cellular ATP?",
    "document": "Mitochondria generate most of a eukaryotic cell's ATP through oxidative phosphorylation."
  },
  {
    "query": "What vulnerability occurs when untrusted input changes a database query?",
    "document": "SQL injection happens when attacker-controlled input alters the structure or meaning of a database query."
  },
  {
    "query": "Why must password hashes use a unique random salt?",
    "document": "A unique salt prevents identical passwords from producing identical hashes and defeats precomputed rainbow tables."
  },
  {
    "query": "Which digest can verify that a downloaded file has not changed?",
    "document": "A published SHA-256 digest can be recomputed and compared to detect accidental or malicious file changes."
  },
  {
    "query": "How can I identify the process listening on a local TCP port on macOS?",
    "document": "The lsof command can map a listening TCP port to its owning process and process identifier."
  },
  {
    "query": "How can separate storage roles appear at stable paths on macOS?",
    "document": "Named APFS volumes mount under stable paths such as /Volumes/Models or /Volumes/Labs."
  },
  {
    "query": "What is a small trainable adapter applied over frozen model weights?",
    "document": "LoRA trains low-rank adapter matrices while leaving the base model weights frozen."
  },
  {
    "query": "What reduces model storage by representing weights with fewer bits?",
    "document": "Weight quantization reduces model size and memory use by storing parameters at lower numerical precision."
  },
  {
    "query": "What calibration artifact records activation importance for enhanced quantization?",
    "document": "An importance matrix, or imatrix, summarizes activation statistics used to guide importance-aware quantization."
  },
  {
    "query": "Which file format commonly packages quantized language models for llama.cpp?",
    "document": "GGUF is a model container format widely used by llama.cpp for quantized weights and metadata."
  },
  {
    "query": "What database supports semantic search using dense numerical representations?",
    "document": "A vector database indexes embedding vectors and retrieves items using similarity measures such as cosine similarity."
  },
  {
    "query": "What SQLite feature supports fast lexical search across document text?",
    "document": "SQLite FTS5 provides an inverted full-text index for efficient token and phrase search."
  },
  {
    "query": "What layer translates a natural-language file question into a database query?",
    "document": "A text-to-SQL layer converts a user's natural-language request into constrained SQL for the catalog."
  },
  {
    "query": "What should be checked before deleting a local backup after transfer?",
    "document": "Verify the destination copy with cryptographic checksums before removing the local source."
  },
  {
    "query": "Where should unique locally trained weights be kept when they are not needed online?",
    "document": "Cold storage is appropriate for unique LoRAs and locally produced model weights that must remain recoverable."
  },
  {
    "query": "What establishes the expected hosts, services, and ports before monitoring changes?",
    "document": "A network baseline records normal hosts, listening services, process owners, and port relationships."
  },
  {
    "query": "How can a file type be recognized when its extension is misleading?",
    "document": "File signatures inspect characteristic bytes in the content rather than trusting only the filename extension."
  },
  {
    "query": "Why does model inference compete with ordinary applications for memory on Apple silicon?",
    "document": "Apple silicon uses unified memory, so CPU workloads, applications, and GPU model tensors share the same physical RAM."
  },
  {
    "query": "Why does a normal Linux Docker container not provide native MLX Metal acceleration?",
    "document": "MLX targets Apple's Metal framework on macOS, while a standard Linux container does not expose that native runtime."
  },
  {
    "query": "Which tool efficiently synchronizes changed files without retransmitting everything?",
    "document": "rsync compares source and destination state and transfers only the data needed to synchronize them."
  },
  {
    "query": "How are normalized embedding vectors commonly ranked for semantic retrieval?",
    "document": "Cosine similarity ranks normalized embeddings by their vector alignment, with larger dot products indicating closer meaning."
  }
]