mnemo-embeddings-bench 0.5.25

Embedding-backend selection benchmark — quality (recall@10, nDCG@10) + latency (p50, p95, vec/s) across OpenAI / ONNX / Noop on a labeled fixture, with an SLA-aware recommender. Anchored on arXiv:2605.23618 (v0.4.9).
Documentation
[
  {"id": "db-01", "topic": "databases", "text": "PostgreSQL uses MVCC to allow concurrent readers without blocking writers."},
  {"id": "db-02", "topic": "databases", "text": "Indexes accelerate row lookup at the cost of extra writes during insert and update."},
  {"id": "db-03", "topic": "databases", "text": "A B-tree index is the default in most relational engines for ordered range scans."},
  {"id": "db-04", "topic": "databases", "text": "Transactions provide ACID guarantees: atomicity, consistency, isolation, durability."},
  {"id": "db-05", "topic": "databases", "text": "DuckDB is a single-file columnar OLAP engine embedded in the host process."},
  {"id": "db-06", "topic": "databases", "text": "Foreign-key constraints enforce referential integrity between related tables."},
  {"id": "db-07", "topic": "databases", "text": "Sharding partitions data across multiple nodes to scale horizontally."},
  {"id": "db-08", "topic": "databases", "text": "Replication keeps secondary nodes in sync with the primary for read scaling and failover."},
  {"id": "db-09", "topic": "databases", "text": "The query planner picks an execution plan based on table statistics and estimated cost."},
  {"id": "db-10", "topic": "databases", "text": "Vacuum reclaims dead tuples left behind by updates and deletes in MVCC engines."},

  {"id": "ml-01", "topic": "machine-learning", "text": "Gradient descent updates model parameters in the direction that reduces the loss."},
  {"id": "ml-02", "topic": "machine-learning", "text": "Cross-entropy loss measures the divergence between predicted and true class distributions."},
  {"id": "ml-03", "topic": "machine-learning", "text": "Regularization techniques like L2 weight decay reduce overfitting on the training set."},
  {"id": "ml-04", "topic": "machine-learning", "text": "Dropout randomly zeroes activations during training to encourage redundant representations."},
  {"id": "ml-05", "topic": "machine-learning", "text": "Batch normalization rescales intermediate activations to stabilize training of deep networks."},
  {"id": "ml-06", "topic": "machine-learning", "text": "Convolutional neural networks share weights across spatial positions for image recognition."},
  {"id": "ml-07", "topic": "machine-learning", "text": "Recurrent networks process sequences one step at a time while maintaining a hidden state."},
  {"id": "ml-08", "topic": "machine-learning", "text": "Transformers replace recurrence with self-attention so every token can look at every other token."},
  {"id": "ml-09", "topic": "machine-learning", "text": "Embedding models map text into dense vectors so semantic similarity becomes cosine similarity."},
  {"id": "ml-10", "topic": "machine-learning", "text": "Fine-tuning adapts a pretrained model to a downstream task with a smaller labeled dataset."},

  {"id": "net-01", "topic": "networking", "text": "TCP provides reliable, ordered, byte-stream delivery between two endpoints."},
  {"id": "net-02", "topic": "networking", "text": "UDP is connectionless and lossy, used when low latency matters more than reliability."},
  {"id": "net-03", "topic": "networking", "text": "The three-way handshake establishes a TCP connection: SYN, SYN-ACK, ACK."},
  {"id": "net-04", "topic": "networking", "text": "TLS encrypts the TCP payload and authenticates the server via X.509 certificates."},
  {"id": "net-05", "topic": "networking", "text": "DNS resolves human-readable hostnames into IP addresses through a hierarchy of nameservers."},
  {"id": "net-06", "topic": "networking", "text": "HTTP/2 multiplexes many request-response pairs over a single TCP connection."},
  {"id": "net-07", "topic": "networking", "text": "QUIC runs over UDP and combines transport, congestion control, and TLS in one protocol."},
  {"id": "net-08", "topic": "networking", "text": "A reverse proxy terminates client connections and forwards traffic to backend servers."},
  {"id": "net-09", "topic": "networking", "text": "Load balancers distribute incoming requests across a pool of healthy backend instances."},
  {"id": "net-10", "topic": "networking", "text": "NAT lets many private IP addresses share a single public IP by rewriting source ports."},

  {"id": "sec-01", "topic": "security", "text": "Public-key cryptography uses a key pair where one key encrypts and the other decrypts."},
  {"id": "sec-02", "topic": "security", "text": "A hash function produces a fixed-size digest that is hard to invert and collision-resistant."},
  {"id": "sec-03", "topic": "security", "text": "HMAC combines a hash function with a secret key to authenticate message integrity."},
  {"id": "sec-04", "topic": "security", "text": "Salted password hashing prevents rainbow-table attacks by making each hash unique."},
  {"id": "sec-05", "topic": "security", "text": "JWT tokens carry signed claims so a server can validate them without a session database."},
  {"id": "sec-06", "topic": "security", "text": "SQL injection happens when user input is concatenated into a query without parameter binding."},
  {"id": "sec-07", "topic": "security", "text": "Cross-site scripting executes attacker JavaScript in a victim's browser through unescaped output."},
  {"id": "sec-08", "topic": "security", "text": "Mutual TLS authenticates both client and server, blocking unauthorized callers at the transport layer."},
  {"id": "sec-09", "topic": "security", "text": "Role-based access control assigns permissions to roles and roles to users."},
  {"id": "sec-10", "topic": "security", "text": "Audit logs record who did what and when so a breach can be investigated after the fact."},

  {"id": "os-01", "topic": "operating-systems", "text": "A process is an isolated execution context with its own virtual address space."},
  {"id": "os-02", "topic": "operating-systems", "text": "Threads within a process share memory and file descriptors but have separate stacks."},
  {"id": "os-03", "topic": "operating-systems", "text": "The scheduler decides which runnable thread to put on each CPU core next."},
  {"id": "os-04", "topic": "operating-systems", "text": "Virtual memory uses page tables to map process addresses to physical RAM or swap."},
  {"id": "os-05", "topic": "operating-systems", "text": "A system call traps from user mode into the kernel to request privileged services."},
  {"id": "os-06", "topic": "operating-systems", "text": "Mutexes serialize access to shared data so only one thread enters the critical section at a time."},
  {"id": "os-07", "topic": "operating-systems", "text": "File descriptors are kernel-side integers that user processes use to identify open files and sockets."},
  {"id": "os-08", "topic": "operating-systems", "text": "Copy-on-write makes fork cheap by sharing pages until one of the processes writes to them."},
  {"id": "os-09", "topic": "operating-systems", "text": "Signals deliver asynchronous notifications to a process, like SIGTERM for graceful shutdown."},
  {"id": "os-10", "topic": "operating-systems", "text": "Cgroups constrain CPU, memory, and I/O for groups of processes on Linux."}
]