Write vectors in fvecs binary format.

Load existing leaderboard (or start empty), add an entry, sort, save, and return.

Allocate a random available port by binding to port 0 and reading the OS-assigned port.

Build the project with `cargo build --release`. Tries an incremental build first. If it fails (e.g. stale/permission-broken artifacts), runs `cargo c

Convert ground truth neighbor IDs to JSON format. Output format per entry: {"query_id": <int>, "neighbors": [<int>, ...]} Args: vect

Compile a tiny Rust program with `rustc` that prints fixed benchmark JSON to stdout and exits 0. Returns the path to the compiled binary.

Create minimal stub data files that the benchmark flow expects.

Discover the base vectors file, handling both single-file and sharded layouts. - If `{data_dir}/base_vectors.json` exists, returns its path directly.

Try to load a previously saved session context. Returns `None` if the file doesn't exist or can't be parsed.

Log an error event.

Log session start.

Get the next round number by scanning existing files in a directory.

Run a correctness test by executing the benchmark with a small query subset. Runs the benchmark client in a lightweight mode and checks whether the r

Run performance profiling on the skeleton server process. Unlike `run_benchmark` and `run_correctness_test`, this function does NOT manage the server

Start the skeleton server binary as a child process. Launches `target/release/<binary>` in `work_dir` with the `PORT` environment variable set to the

Truncate a string to at most `limit` bytes (on a char boundary) and append a notice when truncation occurs.

Wait for a TCP port to become reachable. Polls with `TcpStream::connect` every `poll_interval_ms` milliseconds. Returns `Ok(())` once the port accept

Write vectors in fvecs binary format.

Write vectors in ivecs binary format.

Remove temporary chunk files after successful merge.

Compute KNN for a single query vector using numpy. Returns: List of neighbor IDs (length top_k).

Compute KNN for a single query vector using pure Python. Returns: List of neighbor IDs (length top_k).

Build an assistant message with tool calls (from LLM response).

Build a system message.

Calculate queries per second.

Combine all calculations into a single BenchmarkResult.

Convert query vectors to JSON format. Output format per entry: {"id": <int>, "vector": [<float>, ...]} Args: vectors: List of float

Configurable version of [`detect_hardcoded_results`].

Check whether a path is within the allowed `src/` scope. Normalizes `./`, `\`, and rejects `..` traversal.

Check whether the tool call limit has been reached.

Check whether a path is allowed for reading. Same as `is_allowed_path` but also permits `Cargo.toml`.

Scalar fallback for non-x86_64 platforms

Read a JSON leaderboard file and deserialize into entries.

Load base vectors from a JSON file. The file should contain a JSON array of `{"id": u64, "vector": [f32; 128]}` objects.

Log a tool call (before execution).