Error: Vector queries can fetch at most 256 results, requested 1024.

[apostolisCodpal]

"convex": "^1.23.0"
"@convex-dev/mastra": "0.0.1-alpha.8"
"@mastra/core": "^0.8.3"
"@mastra/memory": "^0.2.10"

When using semantic search (in my test I used OpenAI embedding) I observe the following error:

[CONVEX M(vector/vector:createIndex)] [LOG] 'Creating index memory_messages with dimensions 1536'
[CONVEX A(vector/vector:search)] Uncaught Error: Vector queries can fetch at most 256 results, requested 1024.
    at async handler (../../node_modules/.pnpm/@convex-dev+mastra@0.0.1-alpha.8_@mastra+core@0.8.3_react...)

I believe the issue is coming from the following section in "@convex-dev/mastra":

// src/component/vector/vector.ts
171    const results = await ctx.vectorSearch(index.tableName, "vector", {
172      vector: queryVector,
173      limit: Math.max(topK * 10 * Object.keys(filter ?? {}).length, 1024),
174      filter: filter
175        ? (q) => {
176            return q.eq("indexName", index.indexName);
177          }
178        : undefined,
179    });

I won't pretend to understand the calculation being done for the limit field, but the convex docs clearly say that limit must be between 1 and 256 inclusive.

[ianmacartney]

Good catch. This was an oversight and yeah that calculation is pretty flimsy right now. The goal is to do enough overfetching to post-filter on fields we didn’t include in the query. My get-convex/agent code is much better, as it incorporates specific filters into the vector filter field directly.

…

On Thu, Apr 24, 2025 at 5:06 PM apostolisCodpal ***@***.***> wrote: *apostolisCodpal* created an issue (get-convex/mastra#2) <#2> "convex": "^1.23.0" ***@***.***/mastra": "0.0.1-alpha.8" ***@***.***/core": "^0.8.3" ***@***.***/memory": "^0.2.10" When using semantic search (in my test I used OpenAI embedding) I observe the following error: [CONVEX M(vector/vector:createIndex)] [LOG] 'Creating index memory_messages with dimensions 1536' [CONVEX A(vector/vector:search)] Uncaught Error: Vector queries can fetch at most 256 results, requested 1024. at async handler ***@***.******@***.******@***.******@***.***_react...) I believe the issue is coming from the following section in ***@***.***/mastra": // src/component/vector/vector.ts171 const results = await ctx.vectorSearch(index.tableName, "vector", {172 vector: queryVector,173 limit: Math.max(topK * 10 * Object.keys(filter ?? {}).length, 1024),174 filter: filter175 ? (q) => {176 return q.eq("indexName", index.indexName);177 }178 : undefined,179 }); I won't pretend to understand the calculation being done for the limit field, but the convex docs clearly say that limit must be between 1 and 256 inclusive. — Reply to this email directly, view it on GitHub <#2>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACZQWZTYXNHYZ5HQWJN7JT23D4WVAVCNFSM6AAAAAB3ZJLTJOVHI2DSMVQWIX3LMV43ASLTON2WKOZTGAYTONRQGA4DSMY> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[ianmacartney]

The biggest problem I see right now is that the vector search isn't specifically segmenting by a threadId / userId. It's relying on post-filtering for that. Which is really unlikely to have enough results if you have thousands of threads/users. I'd like to figure out what properties are being passed into the vector search and special-case a few to index on, like I do for the Agent component

[ianmacartney]

I pushed a limit to 256 to main for now though