Kill a server child process and wait for it to exit. Attempts `child.kill()` followed by `child.wait()` to ensure cleanup.

Build extra body fields based on thinking mode setting.

Parse a ToolCall from the function name and JSON arguments string.

Record a completed tool call in the log and increment the counter.

Detect hardcoded / low-diversity results across queries. `results[i]` contains the top-K IDs returned for query `i`. Two checks are performed: 1. **

Return the current session status.

Parse an fvecs file and return a list of float32 vectors. Each vector is stored as: 4 bytes (int32 dimension) + dim * 4 bytes (float32 va

Calculate the p-th percentile from a mutable slice of latencies. Sorts the slice in place and returns the value at the given percentile. `p` should b

Compute ground truth nearest neighbors via brute-force search. This is the in-memory API used by tests and programmatic callers. For large da

Write a log event as a single JSON line, flushed immediately.

Log session end.

Parse perf report --stdio output to extract function names and percentages. Lines look like: " 42.50% 1234 binary_name [.] function_name"

Write content to a file relative to `base_dir`. Only paths within `src/` are allowed. Also checks the readonly list.

Compute squared L2 distance between two vectors (pure Python). We skip the sqrt since it's monotonic — sorting by squared distance gives the

List files in a directory relative to `base_dir`. Only paths within `src/` are allowed. Returns a sorted list of filenames.

Log an LLM response (after receiving).

Return the log file path.

Read a file relative to `base_dir` and return its content. Only paths within `src/` (and `Cargo.toml`) are allowed.

Parse an fvecs file and return a list of float32 vectors. Each vector is stored as: 4 bytes (int32 dimension) + dim * 4 bytes (float32 va

Parse an ivecs file and return a list of int32 vectors. Each vector is stored as: 4 bytes (int32 dimension) + dim * 4 bytes (int32 values

Run the benchmark client against the vector database service. Launches the benchmark binary with the given concurrency and warmup settings, captures

Calculate average recall across all queries. For each query, computes the intersection of the model's top-10 result IDs with the ground truth top-10

Compute the Jaccard similarity between two sets of IDs. Jaccard(A, B) = |A ∩ B| / |A ∪ B|. Returns 0.0 when both sets are empty.

Log an LLM request (before sending).

Execute a shell command inside `base_dir` with an optional timeout. Uses `sh -c` on unix and `cmd /C` on windows. Captures stdout, stderr, exit code

Save session context to `<work_dir>/session_context.json`.

Sort leaderboard entries by QPS descending; ties broken by recall descending.

Write data to a JSON file.

Build an assistant message with text content (no tool calls).

Build a tool result message.

Convert base vectors to JSON format, with optional sharding. Output format per entry: {"id": <int>, "vector": [<float>, ...]} Args:

Scaffold a minimal Cargo project whose binary listens on the PORT env var.

Dispatch a tool call to the appropriate handler. `base_dir` is the project root directory used for file operations, command execution, and benchmark/

Returns JSON Schema tool definitions in OpenAI function calling format.