/**
* postgres-mcp - pgvector Tool Schemas
*
* Input validation schemas for vector similarity search.
* Supports parameter smoothing: col -> column, tableName -> table
*/
import { z } from 'zod';
/**
* Validates that an array contains only finite numbers (rejects Infinity, -Infinity, NaN).
* Provides clear error message instead of confusing "expected number, received number".
*/
export const FiniteNumberArray = z.array(z.number()).superRefine((arr: number[], ctx) => {
const invalidIndexes: number[] = arr
.map((n: number, i: number) => Number.isFinite(n) ? -1 : i)
.filter((i: number) => i >= 0);
if (invalidIndexes.length > 0) {
const invalidValues = invalidIndexes.map((i: number) => String(arr[i])).join(', ');
ctx.addIssue({
code: 'custom',
message: `Vector contains invalid values at index ${invalidIndexes.join(', ')}: ${invalidValues}. Only finite numbers are allowed (no Infinity or NaN).`
});
}
});
// Base schema for MCP exposure (shows all accepted parameters)
const VectorSearchSchemaBase = z.object({
table: z.string().optional().describe('Table name'),
tableName: z.string().optional().describe('Alias for table'),
column: z.string().optional().describe('Vector column name'),
col: z.string().optional().describe('Alias for column'),
vector: FiniteNumberArray.describe('Query vector'),
metric: z.enum(['l2', 'cosine', 'inner_product']).optional().describe('Distance metric'),
limit: z.number().optional().describe('Number of results'),
select: z.array(z.string()).optional().describe('Additional columns to return'),
where: z.string().optional().describe('Filter condition'),
filter: z.string().optional().describe('Alias for where'),
schema: z.string().optional().describe('Database schema (default: public)'),
excludeNull: z.boolean().optional().describe('Exclude rows with NULL vectors (default: false)')
});
// Transformed schema with alias resolution
export const VectorSearchSchema = VectorSearchSchemaBase.transform((data) => ({
table: data.table ?? data.tableName ?? '',
column: data.column ?? data.col ?? '',
vector: data.vector,
metric: data.metric,
limit: data.limit,
select: data.select,
where: data.where ?? data.filter,
schema: data.schema,
excludeNull: data.excludeNull
}));
// Base schema for MCP exposure
const VectorCreateIndexSchemaBase = z.object({
table: z.string().optional().describe('Table name'),
tableName: z.string().optional().describe('Alias for table'),
column: z.string().optional().describe('Vector column name'),
col: z.string().optional().describe('Alias for column'),
type: z.enum(['ivfflat', 'hnsw']).describe('Index type'),
metric: z.enum(['l2', 'cosine', 'inner_product']).optional().describe('Distance metric (default: l2)'),
ifNotExists: z.boolean().optional().describe('Skip if index already exists (default: false)'),
lists: z.number().optional().describe('Number of lists for IVFFlat'),
m: z.number().optional().describe('HNSW m parameter'),
efConstruction: z.number().optional().describe('HNSW ef_construction parameter'),
schema: z.string().optional().describe('Database schema (default: public)')
});
// Transformed schema with alias resolution
export const VectorCreateIndexSchema = VectorCreateIndexSchemaBase.transform((data) => ({
table: data.table ?? data.tableName ?? '',
column: data.column ?? data.col ?? '',
type: data.type,
metric: data.metric ?? 'l2',
ifNotExists: data.ifNotExists,
lists: data.lists,
m: data.m,
efConstruction: data.efConstruction,
schema: data.schema
}));
