Decompress a Parquet GZIP page. The Parquet GZIP codec wraps a raw deflate stream in the standard RFC 1952 gzip envelope; we strip the header (and ign

Composite key hasher matching RadixMultiAggBigKey's BigKeyHash so the shard distribution stays consistent. Mum-style mix of two 64-bit hashes.

Stage 1 inner loop. Per-row IncrementByHashed on (int, str) pairs. Specialized for the gstr × int distinct shape: dict path with precomputed per-dict-

Skip-di variant: SearchPhrase<>'' folded into dict-idx skip.

Two-column inner loop body. Per-RG ingest of (int_key, dict_idx) → count. Precomputes per-dict-entry hash once per RG (~10-30 ns saved per row), then

An identifier "continue" byte: ASCII alnum / underscore, OR any high-bit byte (>= 0x80). Accepting high-bit bytes lets UTF-8 characters (Chinese, Japa

v2: incremental append. When the live view is a simple `SELECT * FROM '<csv>'` and the file grows without the leading bytes changing, parse only the n

SQL keywords, common builtins, and typical aliases we never flag as "not in schema". Conservative: if a token matches any of these, it's definitely no

Virtual view support: stores the original SQL query for re-execution.

Type ID for our simplified format.

Rules that fire on a file-source intent - no schema lookup needed. The file may not even be imported yet; that is the point. Covers the four common sh

Strip stopwords but keep positional tokens that matter for patterns ("how many", "top N", "per", "by", "in"). Preserves order.

Returns false if the footer isn't a valid Thrift FileMetaData.

Parse a PageHeader. Returns bytes consumed.

Wrap multi-word AS aliases in double quotes. Qwen sometimes emits `COUNT(*) AS Total Customers` which the parser rejects because `Customers` starts a

Post-process the model's SQL: wrap any schema identifier that contains a space with double quotes, unless it is already quoted or is inside a string l

Legacy (custom-format) column chunk reader - reads from the mapped buffer.

Standard-Parquet column chunk reader - reads from the mapped buffer.

Render a DDL block for the prompt. Capped at 20 tables - enough for realistic shell sessions, prevents the prompt from exploding if the user has an en

Build the chat-formatted prompt. Qwen2.5 uses the ChatML format (<|im_start|>...<|im_end|>). We render that directly - llama.cpp's chat-template API w

Signal in-flight workers to stop picking new RGs. Workers that have already started decoding the current RG finish it; remaining undecoded RGs are dro

Resolve the DISPLAY logical type for one SELECT-list expression so a Parquet DATE / TIMESTAMP column renders as an ISO string. The engine carries DATE

Rewrite FROM/JOIN references to a file-source alias back to the real path as a single-quoted file literal. Qwen reads the "-- File source" hint in the

Number of row groups (for parallel scheduling).

Parent-projection pushdown: when a PhysicalProjection above this OrderBy only references columns {c1, c2, ...}, pass-2 of CollectTopN_Primitive (which

Reset start_row_ - used when splicing row groups built by parallel workers into a DataTable; the worker builds with start_row_=0 and the splice step r

Per-column hint: decoder writes only str_dict_used bitmap (skip str_dict_indices materialization). Used by the top-N VARCHAR dispatch when we need an

Translate TopN hint's output col_idx through this projection. If expressions_[col_idx] is a simple ColumnRef, we can forward the underlying input col_

Re-tag the logical type WITHOUT touching the backing buffer. Only safe when the new type shares this vector's physical layout (e.g. swapping INTEGER<-

Skip the gzip header (RFC 1952) and return a pointer to the start of the raw deflate stream, or nullptr on malformed header. miniz doesn't support the

--- Snappy decompressor. Calls into vendored google/snappy 1.2.1 (third_party/snappy/) for the SSSE3 16-byte literal-shuffle + IncrementalCopy back-re