turbine-orm 0.16.0 → 0.19.0

package/dist/query/builder.js

@@ -13,7 +13,7 @@
 import { postgresDialect } from '../dialect.js';
 import { CircularRelationError, NotFoundError, OptimisticLockError, RelationError, TimeoutError, ValidationError, wrapPgError, } from '../errors.js';
 import { executeNestedCreate, executeNestedUpdate, hasRelationFields, } from '../nested-write.js';
-import { camelToSnake, snakeToCamel } from '../schema.js';
+import { camelToSnake, normalizeKeyColumns, snakeToCamel } from '../schema.js';
 import { escapeLike, LRUCache, OPERATOR_KEYS, sqlToPreparedName } from './utils.js';
 // ---------------------------------------------------------------------------
 // Internal detection helpers — used by QueryInterface
@@ -120,6 +120,43 @@ function isTextSearchFilter(value) {
 function validateTextSearchConfig(config) {
     return /^[a-zA-Z0-9_]+$/.test(config);
 }
+/**
+ * pgvector distance metric → operator allow-list. This is the ONLY mapping
+ * from a user-supplied metric token to a SQL operator; any token not present
+ * here is rejected, so a user value can never become an arbitrary operator.
+ *
+ *  - `l2`     → `<->` (Euclidean / L2 distance)
+ *  - `cosine` → `<=>` (cosine distance)
+ *  - `ip`     → `<#>` (negative inner product)
+ */
+const VECTOR_METRIC_OPERATORS = {
+    l2: '<->',
+    cosine: '<=>',
+    ip: '<#>',
+};
+/** Comparison keys allowed on a {@link VectorDistanceFilter}. */
+const VECTOR_DISTANCE_COMPARATORS = {
+    lt: '<',
+    lte: '<=',
+    gt: '>',
+    gte: '>=',
+};
+/** Check if a value is a vector distance WHERE filter: `{ distance: { to, metric } }` */
+function isVectorFilter(value) {
+    if (value === null || typeof value !== 'object' || Array.isArray(value) || value instanceof Date) {
+        return false;
+    }
+    const dist = value.distance;
+    return (typeof dist === 'object' &&
+        dist !== null &&
+        !Array.isArray(dist) &&
+        'to' in dist &&
+        'metric' in dist);
+}
+/** Check if an orderBy value is a vector KNN ordering: `{ distance: { to, metric } }` */
+function isVectorOrderBy(value) {
+    return isVectorFilter(value);
+}
 // biome-ignore lint/complexity/noBannedTypes: {} means "no relations known" — intentional for untyped table access
 export class QueryInterface {
     pool;
@@ -151,6 +188,14 @@ export class QueryInterface {
     columnArrayTypeMap;
     /** Tracks tables that have already triggered a deep-with warning (one-time) */
     deepWithWarned = new Set();
+    /**
+     * Per-table memo of date columns keyed by their camelCase FIELD name.
+     * `meta.dateColumns` is keyed by raw snake_case column name, which matches
+     * top-level rows from pg. Nested relation rows arrive from json_build_object
+     * with camelCase keys, so they need this camelCase-keyed set to be coerced
+     * to Date as well (otherwise nested dates leak through as strings).
+     */
+    camelDateFieldCache = new Map();
     /** True when this QI runs inside an active transaction (set via _txScoped option). */
     txScoped;
     /** Original options reference — forwarded to child QIs in nested writes. */
@@ -493,7 +538,16 @@ export class QueryInterface {
         const withFp = args?.with ? this.withFingerprint(args.with) : '';
         const orderFp = args?.orderBy
             ? Object.entries(args.orderBy)
-                .map(([k, d]) => `${k}:${d}`)
+                .map(([k, d]) => {
+                // Vector KNN ordering changes the emitted SQL operator by metric and
+                // adds a `::vector` param, so the metric + direction must be part of
+                // the cache key — otherwise two KNN queries differing only in metric
+                // would collide on a single cached SQL string.
+                if (isVectorOrderBy(d)) {
+                    return `${k}:vec(${d.distance.metric},${d.distance.direction ?? 'asc'})`;
+                }
+                return `${k}:${d}`;
+            })
                 .join(',')
             : '';
         const cursorFp = args?.cursor
@@ -553,7 +607,9 @@ export class QueryInterface {
                 }
             }
             if (args?.orderBy) {
-                sql += ` ORDER BY ${this.buildOrderBy(args.orderBy)}`;
+                // Pass freshParams so vector KNN ordering binds its `$n::vector` query
+                // vector at the correct position (after cursor params, before LIMIT).
+                sql += ` ORDER BY ${this.buildOrderBy(args.orderBy, freshParams)}`;
             }
             if (effectiveLimit !== undefined) {
                 freshParams.push(Number(effectiveLimit));
@@ -581,11 +637,16 @@ export class QueryInterface {
                 params.push(v);
             }
         }
-        // 4. LIMIT param
+        // 4. ORDER BY params (vector KNN ordering binds a `$n::vector` query vector).
+        //    Mirrors buildOrderBy's push order — between cursor and LIMIT.
+        if (args?.orderBy) {
+            this.collectOrderByParams(args.orderBy, params);
+        }
+        // 5. LIMIT param
         if (effectiveLimit !== undefined) {
             params.push(Number(effectiveLimit));
         }
-        // 5. OFFSET param
+        // 6. OFFSET param
         if (args?.offset !== undefined) {
             params.push(Number(args.offset));
         }
@@ -1276,6 +1337,15 @@ export class QueryInterface {
             }
         }
         let sql = `SELECT ${selectExprs.join(', ')} FROM ${this.q(this.table)}${whereSql} GROUP BY ${groupCols.join(', ')}`;
+        // HAVING — filter whole groups by their aggregate values.
+        // Appends to the same `params` array, so placeholders continue from the
+        // WHERE clause's parameter positions (this.p(params.length) below).
+        if (args.having) {
+            const havingClauses = this.buildHavingClauses(args.having, params);
+            if (havingClauses.length > 0) {
+                sql += ` HAVING ${havingClauses.join(' AND ')}`;
+            }
+        }
         // ORDER BY
         if (args.orderBy) {
             sql += ` ORDER BY ${this.buildOrderBy(args.orderBy)}`;
@@ -1343,6 +1413,117 @@ export class QueryInterface {
             tag: `${this.table}.groupBy`,
         };
     }
+    /**
+     * Build the SQL fragments for a {@link HavingClause}.
+     *
+     * Each aggregate expression (`COUNT(*)`, `SUM("col")`, etc.) is constructed
+     * from a **schema-validated, quoted** column identifier — `this.toColumn()`
+     * throws {@link ValidationError} for unknown fields and `this.q()` quotes via
+     * the dialect, so no unvalidated identifier ever reaches the SQL string. Every
+     * comparison value is pushed onto the shared `params` array and referenced by
+     * a `$N` placeholder via {@link buildHavingNumericClauses} — there is no string
+     * interpolation of user values.
+     */
+    buildHavingClauses(having, params) {
+        const clauses = [];
+        // Maps the per-field aggregate key to its SQL function name. The set of
+        // allowed keys is fixed here — any other key on a field's filter object is
+        // rejected by ValidationError below (never interpolated).
+        const aggFnByKey = {
+            _sum: 'SUM',
+            _avg: 'AVG',
+            _min: 'MIN',
+            _max: 'MAX',
+            _count: 'COUNT',
+        };
+        for (const [key, value] of Object.entries(having)) {
+            if (value === undefined)
+                continue;
+            // Top-level `_count` (no field) → COUNT(*) for the whole group.
+            if (key === '_count') {
+                clauses.push(...this.buildHavingNumericClauses('COUNT(*)', value, params));
+                continue;
+            }
+            // Otherwise `key` is a field name mapping to a per-aggregate filter object.
+            if (typeof value !== 'object' || value === null) {
+                throw new ValidationError(`[turbine] Invalid having filter for field "${key}" on table "${this.table}": ` +
+                    `expected an aggregate object like { _sum: { gt: 100 } }.`);
+            }
+            // toColumn validates the field against schema metadata (throws
+            // ValidationError on unknown columns) and q() quotes the identifier — no
+            // unvalidated identifier ever reaches the SQL string.
+            const quotedCol = this.q(this.toColumn(key));
+            for (const [aggKey, filter] of Object.entries(value)) {
+                if (filter === undefined)
+                    continue;
+                const fn = aggFnByKey[aggKey];
+                if (!fn) {
+                    throw new ValidationError(`[turbine] Unknown aggregate "${aggKey}" in having for field "${key}" on table "${this.table}". ` +
+                        `Supported: ${Object.keys(aggFnByKey).join(', ')}.`);
+                }
+                const expr = `${fn}(${quotedCol})`;
+                clauses.push(...this.buildHavingNumericClauses(expr, filter, params));
+            }
+        }
+        return clauses;
+    }
+    /**
+     * Convert a single having filter into one or more parameterized SQL
+     * comparisons against the given aggregate expression. A bare number is
+     * shorthand for equality. Unknown operator keys throw {@link ValidationError}.
+     */
+    buildHavingNumericClauses(expr, filter, params) {
+        // Bare number → equality.
+        if (typeof filter === 'number') {
+            params.push(filter);
+            return [`${expr} = ${this.p(params.length)}`];
+        }
+        if (typeof filter !== 'object' || filter === null) {
+            throw new ValidationError(`[turbine] Invalid having filter on "${expr}" for table "${this.table}": expected a number or operator object.`);
+        }
+        const op = filter;
+        const allowedKeys = new Set(['equals', 'not', 'gt', 'gte', 'lt', 'lte', 'in', 'notIn']);
+        for (const k of Object.keys(op)) {
+            if (!allowedKeys.has(k)) {
+                throw new ValidationError(`[turbine] Unknown having operator "${k}" on "${expr}" for table "${this.table}". ` +
+                    `Supported: ${[...allowedKeys].join(', ')}.`);
+            }
+        }
+        const clauses = [];
+        if (op.equals !== undefined) {
+            params.push(op.equals);
+            clauses.push(`${expr} = ${this.p(params.length)}`);
+        }
+        if (op.not !== undefined) {
+            params.push(op.not);
+            clauses.push(`${expr} != ${this.p(params.length)}`);
+        }
+        if (op.gt !== undefined) {
+            params.push(op.gt);
+            clauses.push(`${expr} > ${this.p(params.length)}`);
+        }
+        if (op.gte !== undefined) {
+            params.push(op.gte);
+            clauses.push(`${expr} >= ${this.p(params.length)}`);
+        }
+        if (op.lt !== undefined) {
+            params.push(op.lt);
+            clauses.push(`${expr} < ${this.p(params.length)}`);
+        }
+        if (op.lte !== undefined) {
+            params.push(op.lte);
+            clauses.push(`${expr} <= ${this.p(params.length)}`);
+        }
+        if (op.in !== undefined) {
+            params.push(op.in);
+            clauses.push(`${expr} = ANY(${this.p(params.length)})`);
+        }
+        if (op.notIn !== undefined) {
+            params.push(op.notIn);
+            clauses.push(`${expr} != ALL(${this.p(params.length)})`);
+        }
+        return clauses;
+    }
     // -------------------------------------------------------------------------
     // aggregate — standalone aggregation without groupBy
     // -------------------------------------------------------------------------
@@ -1506,12 +1687,24 @@ export class QueryInterface {
      */
     resolveColumns(select, omit) {
         if (select) {
+            // An array here means a caller wrote `select: ['id', 'name']` (Drizzle/SQL
+            // style) instead of the object shape. Object.entries() would iterate the
+            // numeric indices and throw a cryptic `Unknown field "0"` — catch it early
+            // with an actionable message.
+            if (Array.isArray(select)) {
+                throw new ValidationError(`[turbine] "select" must be an object mapping field names to true ` +
+                    `(e.g. { id: true, name: true }), not an array.`);
+            }
             // Only include columns where value is true
             return Object.entries(select)
                 .filter(([, v]) => v)
                 .map(([k]) => this.toColumn(k));
         }
         if (omit) {
+            if (Array.isArray(omit)) {
+                throw new ValidationError(`[turbine] "omit" must be an object mapping field names to true ` +
+                    `(e.g. { createdAt: true }), not an array.`);
+            }
             // Include all columns except those where value is true
             const omitCols = new Set(Object.entries(omit)
                 .filter(([, v]) => v)
@@ -1687,6 +1880,17 @@ export class QueryInterface {
                 parts.push(`${key}:op(${opKeys.join(',')}${modeStr})`);
                 continue;
             }
+            // Vector distance filter — metric (operator) and present comparators
+            // change the SQL shape, so both go in the fingerprint.
+            if (typeof value === 'object' && !Array.isArray(value) && isVectorFilter(value)) {
+                const dist = value.distance;
+                const cmps = Object.keys(VECTOR_DISTANCE_COMPARATORS)
+                    .filter((c) => dist[c] !== undefined)
+                    .sort()
+                    .join('|');
+                parts.push(`${key}:vec(${dist.metric},${cmps})`);
+                continue;
+            }
             // JSON filter
             if (typeof value === 'object' && !Array.isArray(value) && isJsonFilter(value)) {
                 const jKeys = Object.keys(value).sort();
@@ -1811,6 +2015,14 @@ export class QueryInterface {
             if (value === null)
                 continue;
             const rawColumn = this.toColumn(key);
+            // Vector distance filter — mirrors buildVectorFilterClauses push order.
+            if (typeof value === 'object' && !Array.isArray(value) && isVectorFilter(value)) {
+                // Validate the same way the build path does so the collect path never
+                // diverges (it would throw before any param was pushed).
+                this.vectorOperator(key, rawColumn, value.distance.metric);
+                this.collectVectorFilterParams(key, rawColumn, value, params);
+                continue;
+            }
             // JSONB filter
             if (typeof value === 'object' && !Array.isArray(value) && isJsonFilter(value)) {
                 const colType = this.getColumnPgType(rawColumn);
@@ -1907,6 +2119,37 @@ export class QueryInterface {
             params.push(filter.hasSome);
         // isEmpty has no params (IS NULL / IS NOT NULL)
     }
+    /**
+     * Collect params for an orderBy clause. Only vector KNN ordering pushes a
+     * param (the `$n::vector` query vector); plain direction ordering is
+     * parameterless. Mirrors buildOrderBy's push order exactly so the cached-SQL
+     * param re-collection stays in lockstep.
+     */
+    collectOrderByParams(orderBy, params) {
+        for (const [key, dir] of Object.entries(orderBy)) {
+            if (isVectorOrderBy(dir)) {
+                const rawColumn = this.toColumn(key);
+                // Re-run the same validation as buildOrderBy so the collect path can
+                // never push a param that the build path rejected (or vice versa).
+                this.vectorOperator(key, rawColumn, dir.distance.metric);
+                this.pushVectorParam(key, rawColumn, dir.distance.to, params);
+            }
+        }
+    }
+    /**
+     * Collect params for a vector distance WHERE filter. Mirrors
+     * {@link buildVectorFilterClauses}: the `$n::vector` query vector first, then
+     * the comparison threshold(s).
+     */
+    collectVectorFilterParams(field, rawColumn, filter, params) {
+        const dist = filter.distance;
+        this.pushVectorParam(field, rawColumn, dist.to, params);
+        for (const cmp of Object.keys(VECTOR_DISTANCE_COMPARATORS)) {
+            const threshold = dist[cmp];
+            if (threshold !== undefined)
+                params.push(threshold);
+        }
+    }
     /**
      * Produce a fingerprint for a `with` clause tree. Recursion mirrors
      * buildSelectWithRelations / buildRelationSubquery.
@@ -2003,6 +2246,27 @@ export class QueryInterface {
         const targetMeta = this.schema.tables[targetTable];
         if (!targetMeta)
             return;
+        // manyToMany param order mirrors buildManyToManySubquery:
+        //   where params → limit param → nested-with params (always, both paths).
+        if (relDef.type === 'manyToMany') {
+            if (spec.where) {
+                for (const [, v] of Object.entries(spec.where)) {
+                    params.push(v);
+                }
+            }
+            if (spec.limit) {
+                params.push(Number(spec.limit));
+            }
+            if (spec.with) {
+                for (const [nestedRelName, nestedSpec] of Object.entries(spec.with)) {
+                    const nestedRelDef = targetMeta.relations[nestedRelName];
+                    if (!nestedRelDef)
+                        continue;
+                    this.collectRelationSubqueryParams(nestedRelDef, nestedSpec, params, 'alias', depth + 1);
+                }
+            }
+            return;
+        }
         const willWrap = relDef.type === 'hasMany' && (spec.limit !== undefined || spec.orderBy !== undefined);
         // Non-wrapped path: nested relations BEFORE where/limit
         if (!willWrap && spec.with) {
@@ -2019,8 +2283,10 @@ export class QueryInterface {
                 params.push(v);
             }
         }
-        // limit param
-        if (spec.limit) {
+        // limit param — only hasMany parameterizes its limit (mirrors
+        // buildRelationSubquery). belongsTo/hasOne ignore limit (always LIMIT 1), so
+        // pushing one here would orphan a param and desync the collect path.
+        if (relDef.type === 'hasMany' && spec.limit) {
             params.push(Number(spec.limit));
         }
         // Wrapped path: nested relations AFTER where/limit (inside inner subquery)
@@ -2174,6 +2440,12 @@ export class QueryInterface {
                 andClauses.push(`${column} IS NULL`);
                 continue;
             }
+            // Handle vector distance filter (pgvector): `{ distance: { to, metric, lt } }`
+            if (typeof value === 'object' && !Array.isArray(value) && isVectorFilter(value)) {
+                const vecClauses = this.buildVectorFilterClauses(key, rawColumn, value, params);
+                andClauses.push(...vecClauses);
+                continue;
+            }
             // Handle JSONB filter operators (for json/jsonb columns)
             if (typeof value === 'object' && !Array.isArray(value) && isJsonFilter(value)) {
                 const colType = this.getColumnPgType(rawColumn);
@@ -2222,6 +2494,25 @@ export class QueryInterface {
                 andClauses.push(...opClauses);
                 continue;
             }
+            // Strict validation: a plain (non-array, non-Date) object on a non-JSON
+            // column matched no known filter shape — almost always a misspelled
+            // operator (`startWith` for `startsWith`) or a stray nested object.
+            // Silently treating it as `col = $1` returns wrong rows with no error, so
+            // throw with the offending keys and the supported operator list. JSON/JSONB
+            // columns legitimately accept object values for equality, so they fall
+            // through unchanged.
+            if (typeof value === 'object' && !Array.isArray(value) && !(value instanceof Date)) {
+                const colType = this.getColumnPgType(rawColumn);
+                if (colType !== 'json' && colType !== 'jsonb') {
+                    const badKeys = Object.keys(value);
+                    throw new ValidationError(badKeys.length === 0
+                        ? `[turbine] Empty filter object on "${rawColumn}" for table "${this.table}". ` +
+                            `Provide a value or an operator like { gt: 1 }.`
+                        : `[turbine] Unknown operator${badKeys.length > 1 ? 's' : ''} ` +
+                            `${badKeys.map((k) => `"${k}"`).join(', ')} on "${rawColumn}" for table "${this.table}". ` +
+                            `Supported operators: ${[...OPERATOR_KEYS].join(', ')}.`);
+                }
+            }
             // Plain equality
             params.push(value);
             andClauses.push(`${column} = ${this.p(params.length)}`);
@@ -2380,8 +2671,17 @@ export class QueryInterface {
         }
         return clauses;
     }
-    /** Build ORDER BY clause from an object */
-    buildOrderBy(orderBy) {
+    /**
+     * Build ORDER BY clause from an object.
+     *
+     * Each value is either a plain direction (`'asc'`/`'desc'`) or — for pgvector
+     * columns — a `{ distance: { to, metric, direction? } }` KNN ordering object.
+     * Vector ordering binds the query vector as a `$n::vector` param, so a `params`
+     * array MUST be supplied when a vector ordering may be present (top-level
+     * findMany path). When `params` is omitted (groupBy / relation path) a vector
+     * ordering throws — KNN ordering is only supported at the top level.
+     */
+    buildOrderBy(orderBy, params) {
         // Dev-only: validate that orderBy fields exist in the table schema
         if (process.env.NODE_ENV !== 'production') {
             for (const key of Object.keys(orderBy)) {
@@ -2396,14 +2696,88 @@ export class QueryInterface {
         return Object.entries(orderBy)
             .map(([key, dir]) => {
             if (meta && !(key in meta.columnMap)) {
-                throw new ValidationError(`Unknown column "${key}" in orderBy for table "${this.table}"`);
+                throw new ValidationError(`[turbine] Unknown field "${key}" in orderBy on table "${this.table}". ` +
+                    `Known fields: ${Object.keys(meta.columnMap).join(', ') || '(none)'}.`);
+            }
+            // Vector KNN ordering: { distance: { to, metric, direction? } }
+            if (isVectorOrderBy(dir)) {
+                if (!params) {
+                    throw new ValidationError(`[turbine] Vector distance ordering on "${key}" is only supported in a top-level findMany orderBy.`);
+                }
+                const rawColumn = this.toColumn(key);
+                const operator = this.vectorOperator(key, rawColumn, dir.distance.metric);
+                const placeholder = this.pushVectorParam(key, rawColumn, dir.distance.to, params);
+                const safeDir = dir.distance.direction?.toLowerCase() === 'desc' ? 'DESC' : 'ASC';
+                return `${this.q(rawColumn)} ${operator} ${placeholder} ${safeDir}`;
             }
             const safeDir = dir.toLowerCase() === 'desc' ? 'DESC' : 'ASC';
             return `${this.toSqlColumn(key)} ${safeDir}`;
         })
             .join(', ');
     }
+    // -------------------------------------------------------------------------
+    // pgvector helpers (similarity search)
+    // -------------------------------------------------------------------------
+    /**
+     * Resolve a {@link VectorMetric} to its pgvector distance operator from a
+     * fixed allow-list, validating the target column is actually a `vector`
+     * column. Throws {@link ValidationError} for an unknown metric or a
+     * non-vector column — a user-supplied string can never become a SQL operator.
+     */
+    vectorOperator(field, rawColumn, metric) {
+        const colType = this.getColumnPgType(rawColumn);
+        if (colType !== 'vector') {
+            throw new ValidationError(`[turbine] Column "${field}" on table "${this.table}" is not a vector column ` +
+                `(actual type: ${colType}); cannot apply a vector distance operation.`);
+        }
+        const op = VECTOR_METRIC_OPERATORS[metric];
+        if (!op) {
+            throw new ValidationError(`[turbine] Unknown vector metric "${metric}" for column "${field}". ` +
+                `Valid metrics: ${Object.keys(VECTOR_METRIC_OPERATORS).join(', ')}.`);
+        }
+        return op;
+    }
+    /**
+     * Validate and bind a query vector as a single `$n::vector` parameter.
+     * Every element must be a finite number (no NaN / Infinity / strings) so a
+     * malformed array can never produce a broken `::vector` literal, and the array
+     * is NEVER string-interpolated into the SQL text. Returns the `$n::vector`
+     * placeholder string.
+     */
+    pushVectorParam(field, _rawColumn, to, params) {
+        if (!Array.isArray(to) || to.length === 0) {
+            throw new ValidationError(`[turbine] Vector distance on "${field}" requires a non-empty array of numbers for "to".`);
+        }
+        for (const el of to) {
+            if (typeof el !== 'number' || !Number.isFinite(el)) {
+                throw new ValidationError(`[turbine] Vector "to" for column "${field}" must contain only finite numbers; ` +
+                    `got ${JSON.stringify(el)}.`);
+            }
+        }
+        // Bind as a pgvector text literal '[1,2,3]'. Elements are already validated
+        // as finite numbers, so the joined string is safe; it is still passed as a
+        // bound param (never interpolated) and cast with ::vector.
+        params.push(`[${to.join(',')}]`);
+        return `${this.p(params.length)}::vector`;
+    }
     /** Parse a flat row: convert snake_case to camelCase + Date coercion */
+    /**
+     * Returns the set of camelCase field names for a table's date columns,
+     * derived once from `meta.dateColumns` (snake_case) via reverseColumnMap and
+     * memoized per table. Used so nested relation rows (camelCase keys) coerce
+     * dates the same way top-level rows do.
+     */
+    getCamelDateFields(table, meta) {
+        let camel = this.camelDateFieldCache.get(table);
+        if (!camel) {
+            camel = new Set();
+            for (const col of meta.dateColumns) {
+                camel.add(meta.reverseColumnMap[col] ?? col);
+            }
+            this.camelDateFieldCache.set(table, camel);
+        }
+        return camel;
+    }
     parseRow(row, table) {
         const parsed = {};
         const meta = this.schema.tables[table];
@@ -2411,12 +2785,16 @@ export class QueryInterface {
             // Fast path: use pre-computed maps (avoids regex per column per row)
             const reverseMap = meta.reverseColumnMap;
             const dateCols = meta.dateColumns;
+            // camelCase-keyed date fields, so nested json_build_object rows (whose
+            // keys are already camelCase) get the same Date coercion as top-level rows.
+            const camelDateFields = this.getCamelDateFields(table, meta);
             const keys = Object.keys(row);
             for (let i = 0; i < keys.length; i++) {
                 const col = keys[i];
                 const value = row[col];
                 const field = reverseMap[col] ?? col; // fall back to raw col name, not regex
-                if (dateCols.has(col) && value !== null && !(value instanceof Date)) {
+                // Top-level rows are snake_case (dateCols); nested rows are camelCase (camelDateFields).
+                if ((dateCols.has(col) || camelDateFields.has(field)) && value !== null && !(value instanceof Date)) {
                     parsed[field] = new Date(value);
                 }
                 else {
@@ -2460,14 +2838,15 @@ export class QueryInterface {
             if (typeof rawValue === 'string') {
                 try {
                     const jsonVal = JSON.parse(rawValue);
-                    // After parsing, apply parseRow to each item for snake→camel + date coercion
+                    // After parsing, recurse via parseNestedRow so each item gets date
+                    // coercion AND its own sub-relations parsed at arbitrary depth.
                     if (Array.isArray(jsonVal)) {
                         parsed[relName] = jsonVal.map((item) => typeof item === 'object' && item !== null
-                            ? this.parseRow(item, relDef.to)
+                            ? this.parseNestedRow(item, relDef.to)
                             : item);
                     }
                     else if (typeof jsonVal === 'object' && jsonVal !== null) {
-                        parsed[relName] = this.parseRow(jsonVal, relDef.to);
+                        parsed[relName] = this.parseNestedRow(jsonVal, relDef.to);
                     }
                     else {
                         parsed[relName] = jsonVal;
@@ -2479,10 +2858,12 @@ export class QueryInterface {
                 }
             }
             else if (Array.isArray(rawValue)) {
-                parsed[relName] = rawValue.map((item) => typeof item === 'object' && item !== null ? this.parseRow(item, relDef.to) : item);
+                parsed[relName] = rawValue.map((item) => typeof item === 'object' && item !== null
+                    ? this.parseNestedRow(item, relDef.to)
+                    : item);
             }
             else if (typeof rawValue === 'object' && rawValue !== null) {
-                parsed[relName] = this.parseRow(rawValue, relDef.to);
+                parsed[relName] = this.parseNestedRow(rawValue, relDef.to);
             }
             else {
                 parsed[relName] = rawValue;
@@ -2683,6 +3064,12 @@ export class QueryInterface {
         // When wrapping, nested relations are built in the wrapped path referencing innerAlias,
         // so we must NOT build them here (they would push orphaned params).
         const willWrap = relDef.type === 'hasMany' && spec !== true && (spec.limit !== undefined || spec.orderBy !== undefined);
+        // manyToMany takes a dedicated JOIN-through-junction path. Nested relations,
+        // where, orderBy, and select/omit are handled there (the target alias is the
+        // row source, exactly like hasMany), so short-circuit before the hasMany logic.
+        if (relDef.type === 'manyToMany') {
+            return this.buildManyToManySubquery(relDef, spec, params, parentRef, aliasCounter, currentDepth, currentPath, alias, targetMeta, targetColumns);
+        }
         // Nested relations — only in the non-wrapped path (wrapped path builds them separately)
         if (!willWrap && spec !== true && spec.with) {
             for (const [nestedRelName, nestedSpec] of Object.entries(spec.with)) {
@@ -2739,9 +3126,13 @@ export class QueryInterface {
                 whereClause += ` AND ${alias}.${this.q(col)} = ${this.p(params.length)}`;
             }
         }
-        // LIMIT
+        // LIMIT — only meaningful for hasMany. A belongsTo / hasOne subquery returns
+        // a single row (literal `LIMIT 1` below), so a `spec.limit` here must NOT push
+        // a parameter: doing so orphans an untyped `$N` that the SQL never references,
+        // which Postgres rejects with "could not determine data type of parameter $N"
+        // (and shifts every later placeholder by one). To-one relations ignore limit.
         let limitClause = '';
-        if (spec !== true && spec.limit) {
+        if (relDef.type === 'hasMany' && spec !== true && spec.limit) {
             params.push(Number(spec.limit));
             limitClause = ` LIMIT ${this.p(params.length)}`;
         }
@@ -2779,6 +3170,146 @@ export class QueryInterface {
         // belongsTo / hasOne — return single object
         return `SELECT ${jsonObj} FROM ${qTarget} ${alias} WHERE ${whereClause} LIMIT 1`;
     }
+    /**
+     * Build the json_agg subquery for a `manyToMany` relation, JOINing the target
+     * table through a junction (join) table.
+     *
+     * Shape (no LIMIT/ORDER):
+     * ```sql
+     * SELECT COALESCE(json_agg(json_build_object(...)), '[]'::json)
+     * FROM <target> <talias>
+     * JOIN <junction> <jalias> ON <jalias>.<targetKey> = <talias>.<targetPK>
+     * WHERE <jalias>.<sourceKey> = <parentRef>.<referenceKey>
+     * ```
+     *
+     * With LIMIT/ORDER, the rows are wrapped in an inner subquery so the LIMIT
+     * applies BEFORE aggregation (identical strategy to hasMany).
+     *
+     * Cardinality is always 'many' → empty-array fallback, never NULL.
+     *
+     * IMPORTANT: every `params.push` here MUST be mirrored, in the same order, in
+     * {@link collectRelationSubqueryParams} or pipeline batching will desync.
+     */
+    buildManyToManySubquery(relDef, spec, params, parentRef, aliasCounter, currentDepth, currentPath, talias, targetMeta, targetColumns) {
+        if (!relDef.through) {
+            throw new ValidationError(`[turbine] manyToMany relation "${relDef.name}" is missing a \`through\` junction descriptor.`);
+        }
+        const targetTable = relDef.to;
+        const qTarget = this.q(targetTable);
+        const qJunction = this.q(relDef.through.table);
+        const qParent = this.q(parentRef);
+        const jalias = `${talias}j`; // junction alias, distinct from the target alias
+        // JOIN: junction.targetKey = target.<targetPK>. Composite keys pair positionally.
+        const targetKeys = normalizeKeyColumns(relDef.through.targetKey);
+        // The target PK is the column(s) the junction's targetKey references. An empty
+        // introspected PK means we cannot know what to JOIN on — fail loudly rather than
+        // silently guessing `id` and generating a wrong JOIN.
+        if (targetMeta.primaryKey.length === 0) {
+            throw new ValidationError(`[turbine] manyToMany relation "${relDef.name}" targets table "${targetTable}" which has no primary key; ` +
+                `cannot determine the join column. Define a primary key or use an explicit through descriptor.`);
+        }
+        const targetPk = targetMeta.primaryKey;
+        if (targetKeys.length !== targetPk.length) {
+            throw new ValidationError(`[turbine] manyToMany relation "${relDef.name}": through.targetKey has ${targetKeys.length} column(s) ` +
+                `but target "${targetTable}" primary key has ${targetPk.length}. Composite keys must pair positionally.`);
+        }
+        const joinOn = targetKeys
+            .map((jcol, i) => `${jalias}.${this.q(jcol)} = ${talias}.${this.q(targetPk[i])}`)
+            .join(' AND ');
+        // Correlation: junction.sourceKey = parent.<referenceKey>.
+        const sourceKeys = normalizeKeyColumns(relDef.through.sourceKey);
+        const refKeys = normalizeKeyColumns(relDef.referenceKey);
+        if (sourceKeys.length !== refKeys.length) {
+            throw new ValidationError(`[turbine] manyToMany relation "${relDef.name}": through.sourceKey has ${sourceKeys.length} column(s) ` +
+                `but referenceKey has ${refKeys.length}. Composite keys must pair positionally.`);
+        }
+        let whereClause = sourceKeys
+            .map((jcol, i) => `${jalias}.${this.q(jcol)} = ${qParent}.${this.q(refKeys[i])}`)
+            .join(' AND ');
+        // ORDER BY on the target rows
+        let orderClause = '';
+        if (spec !== true && spec.orderBy) {
+            const orders = Object.entries(spec.orderBy)
+                .map(([k, dir]) => {
+                const col = camelToSnake(k);
+                if (!targetMeta.allColumns.includes(col)) {
+                    throw new ValidationError(`[turbine] Unknown column "${k}" in orderBy for table "${targetTable}"`);
+                }
+                const safeDir = dir.toLowerCase() === 'desc' ? 'DESC' : 'ASC';
+                return `${talias}.${this.q(col)} ${safeDir}`;
+            })
+                .join(', ');
+            orderClause = ` ORDER BY ${orders}`;
+        }
+        // Additional WHERE filters on the target — properly parameterized.
+        if (spec !== true && spec.where) {
+            for (const [k, v] of Object.entries(spec.where)) {
+                const col = camelToSnake(k);
+                if (!targetMeta.allColumns.includes(col)) {
+                    throw new ValidationError(`[turbine] Unknown column "${k}" in where for table "${targetTable}"`);
+                }
+                params.push(v);
+                whereClause += ` AND ${talias}.${this.q(col)} = ${this.p(params.length)}`;
+            }
+        }
+        // LIMIT
+        let limitClause = '';
+        if (spec !== true && spec.limit) {
+            params.push(Number(spec.limit));
+            limitClause = ` LIMIT ${this.p(params.length)}`;
+        }
+        const fromJoin = `FROM ${qTarget} ${talias} JOIN ${qJunction} ${jalias} ON ${joinOn}`;
+        // When LIMIT or ORDER BY is present, wrap the joined rows in an inner subquery
+        // so the LIMIT applies to rows BEFORE aggregation (same approach as hasMany).
+        if (limitClause || orderClause) {
+            const innerAlias = `${talias}i`;
+            const innerSql = `SELECT ${targetMeta.allColumns.map((c) => `${talias}.${this.q(c)}`).join(', ')} ` +
+                `${fromJoin} WHERE ${whereClause}${orderClause}${limitClause}`;
+            const innerJsonPairs = targetColumns.map((col) => [
+                targetMeta.reverseColumnMap[col] ?? snakeToCamel(col),
+                `${innerAlias}.${this.q(col)}`,
+            ]);
+            // Nested relations reference the inner alias.
+            if (spec !== true && spec.with) {
+                for (const [nestedRelName, nestedSpec] of Object.entries(spec.with)) {
+                    const nestedRelDef = targetMeta.relations[nestedRelName];
+                    if (!nestedRelDef) {
+                        throw new RelationError(`[turbine] Unknown relation "${nestedRelName}" on table "${targetTable}". ` +
+                            `Available: ${Object.keys(targetMeta.relations).join(', ')}`);
+                    }
+                    const nestedSub = this.buildRelationSubquery(nestedRelDef, nestedSpec, params, innerAlias, aliasCounter, currentDepth + 1, [...currentPath, relDef.name]);
+                    const fallback = nestedRelDef.type === 'belongsTo' || nestedRelDef.type === 'hasOne'
+                        ? this.dialect.nullJsonLiteral
+                        : this.dialect.emptyJsonArrayLiteral;
+                    innerJsonPairs.push([nestedRelName, `COALESCE((${nestedSub}), ${fallback})`]);
+                }
+            }
+            const innerJsonObj = this.dialect.buildJsonObject(innerJsonPairs);
+            return `SELECT ${this.dialect.buildJsonArrayAgg(innerJsonObj)} FROM (${innerSql}) ${innerAlias}`;
+        }
+        // Simple path: build the json object pairs directly off the target alias,
+        // including any nested relations (correlated to the target alias).
+        const jsonPairs = targetColumns.map((col) => [
+            targetMeta.reverseColumnMap[col] ?? snakeToCamel(col),
+            `${talias}.${this.q(col)}`,
+        ]);
+        if (spec !== true && spec.with) {
+            for (const [nestedRelName, nestedSpec] of Object.entries(spec.with)) {
+                const nestedRelDef = targetMeta.relations[nestedRelName];
+                if (!nestedRelDef) {
+                    throw new RelationError(`[turbine] Unknown relation "${nestedRelName}" on table "${targetTable}". ` +
+                        `Available: ${Object.keys(targetMeta.relations).join(', ')}`);
+                }
+                const nestedSub = this.buildRelationSubquery(nestedRelDef, nestedSpec, params, talias, aliasCounter, currentDepth + 1, [...currentPath, relDef.name]);
+                const fallback = nestedRelDef.type === 'belongsTo' || nestedRelDef.type === 'hasOne'
+                    ? this.dialect.nullJsonLiteral
+                    : this.dialect.emptyJsonArrayLiteral;
+                jsonPairs.push([nestedRelName, `COALESCE((${nestedSub}), ${fallback})`]);
+            }
+        }
+        const jsonObj = this.dialect.buildJsonObject(jsonPairs);
+        return `SELECT ${this.dialect.buildJsonArrayAgg(jsonObj)} ${fromJoin} WHERE ${whereClause}`;
+    }
     /**
      * Get the Postgres type for a column (e.g. 'jsonb', 'text', '_int4').
      * Used to detect JSONB/array columns for specialized operators.
@@ -2872,6 +3403,39 @@ export class QueryInterface {
         }
         return clauses;
     }
+    /**
+     * Build SQL clauses for a pgvector distance WHERE filter:
+     *
+     *   `"embedding" <-> $1::vector < $2`
+     *
+     * The query vector is bound as a `$n::vector` param (never interpolated), the
+     * metric maps to an operator via a fixed allow-list, and each comparison
+     * threshold (`lt`/`lte`/`gt`/`gte`) is its own bound param. Emits one clause
+     * per supplied comparator (all ANDed). Param push order matches
+     * {@link collectVectorFilterParams}.
+     */
+    buildVectorFilterClauses(field, rawColumn, filter, params) {
+        const dist = filter.distance;
+        const operator = this.vectorOperator(field, rawColumn, dist.metric);
+        const placeholder = this.pushVectorParam(field, rawColumn, dist.to, params);
+        const distanceExpr = `${this.q(rawColumn)} ${operator} ${placeholder}`;
+        const clauses = [];
+        for (const [cmp, sqlOp] of Object.entries(VECTOR_DISTANCE_COMPARATORS)) {
+            const threshold = dist[cmp];
+            if (threshold === undefined)
+                continue;
+            if (typeof threshold !== 'number' || !Number.isFinite(threshold)) {
+                throw new ValidationError(`[turbine] Vector distance threshold "${cmp}" on "${field}" must be a finite number; ` +
+                    `got ${JSON.stringify(threshold)}.`);
+            }
+            params.push(threshold);
+            clauses.push(`${distanceExpr} ${sqlOp} ${this.p(params.length)}`);
+        }
+        if (clauses.length === 0) {
+            throw new ValidationError(`[turbine] Vector distance filter on "${field}" requires at least one comparison (lt / lte / gt / gte).`);
+        }
+        return clauses;
+    }
     /**
      * Build SQL clause for full-text search using to_tsvector @@ to_tsquery.
      * The config name is validated to prevent injection (only alphanumeric + underscore).