Ingest one VARCHAR value into thread `tid`'s bounded active table.

Field accessors mirror old direct access (so reads on never-used states return defaults without allocation).

* Return the underlying SlothDB engine version string. * * @returns {string}

Attach an externally-owned contiguous byte buffer. The buffer is kept alive for the Vector's lifetime. Used when another decoder (e.g. Parquet) alread

Map an Avro primitive type name + optional logicalType to a SlothDB LogicalType. Logical types we recognise: - timestamp-millis / timestamp-micros → T

Structural equality for two bound-expression trees. Used to detect when a SELECT-list / ORDER-BY subtree matches a GROUP BY entry verbatim so we can r

Coerce `v` to match the type of `target` for cross-type comparisons. Value::operator< returns false on type mismatch, so without this step `BIGINT(200

SUM fast path: walk pages of a BIGINT dict-encoded column, accumulate histogram of dict indices, and reduce SUM = sum_idx (count[idx]*dict[idx]). Skip

Reverse of EncodeCell: rebuild a typed Value from the binary blob.

Decode a DataPage (V1 or V2) into `out` values. Handles PLAIN and PLAIN_DICTIONARY. For flat schemas with nullable columns, def_level bit_width is 1.

Decode one data page (V1 or V2) into the typed output buffer starting at row_offset. Supports PLAIN and PLAIN_DICTIONARY / RLE_DICTIONARY. `skip_str_d

PLAIN BOOLEAN is bit-packed (1 bit per value).

Append PLAIN-encoded VARCHAR data into `heap`, writing string_t entries to dst[i] for non-null rows. heap is a pre-reserved buffer (no reallocation).

Mask-aware variant: skip dst[i] writes for rows where filter_mask[i]==0. The 4-byte length read still happens (PLAIN has no random access — we must ad

Lengths-only PLAIN VARCHAR decode. Reads the 4-byte length prefix per row into `dst_lengths[i]`, then advances past the byte data without copying. Use

2-stage COUNT(DISTINCT INT) GROUP BY VARCHAR: Phase A (parallel) builds per-shard local str_count maps from the unique-pair set; Phase B reduces seque

Encode one group-column cell into `key` as a length-self-describing binary blob: a 1-byte presence tag, then a fixed- or length-prefixed payload deter

Extract `as <name>` or `into <name>` after `after_pos`. Returns the raw identifier (no quoting) or empty if absent. Also accepts `named X` and `called

Extract the "X by|per|grouped-by Y" group column token, if any. Returns the noun token or empty.

Rewrite bare identifiers in the SQL that fuzzily match a schema name but aren't the canonical form. Catches Qwen's common mistakes: `customer_id` -> `

Extract an integer literal from a value expression. Accepts a plain CONSTANT and a UNARY_MINUS wrapping one — the parser turns `-1` into UnaryMinus(Co

Try to extract simple `column OP literal` predicates from a bound expression. Walks AND-chains and accepts =, <, <=, >, >= on a column vs. a constant.

Strip markdown fences + trim + drop trailing semicolons from model output.

Extract a "top N" limit from tokens. Returns 0 if absent.

Extract a 4-digit year from tokens after "in"/"during"/"for" - very common NL pattern ("sales in 2024"). Returns empty if absent.

Find the position of a specific keyword in a token list.

Cross-check every identifier in the generated SQL against the schema. Returns a list of column-shaped names that don't match any table or column. Used

Format a TIMESTAMP value (microseconds since 1970-01-01) as "YYYY-MM-DD HH:MM:SS[.uuuuuu]". Avoids gmtime so this is reentrant and works the same on e

--- Fuse interface: used by PhysicalHashAggregate to bypass materialization. ---