opensearch-project
[BUG]Agentic Search <-> Ollama serialization error
Originally from https://github.com/opensearch-project/OpenSearch/issues/20402#issue-3801630102
I’ve followed all the steps to configure agentic search (conversational agent) on my local OpenSearch server to use a local LLM running under Ollama. However, my query produces the error below. See the HTTP traces showing that a property is incorrectly serialized as a float instead of an int.
This is the query I issued using curl (sorry for the formatting):
curl -k -u admin:admin -X GET “``http://localhost:9200/able_chunks_v1/_search?search_pipeline=agentic-pipeline``” -H “Content-Type: application/json” -d ‘{
“query”: {
“agentic”: {
“query_text”: “How many documents are there in the index”
}
}
}’
And this is the error: ”json: cannot unmarshal number 0.0 into Go struct field ToolCall.messages.tool_calls.index of type int”
Full error:
{“error”:{“root_cause”:[{“type”:“illegal_argument_exception”,“reason”:“Agentic search failed - Agent execution error - Agent ID: [_Nh7rZsBMCptIK-aGFFT], Error: [Error from remote service: {"error":{"message":"json: cannot unmarshal number 0.0 into Go struct field ToolCall.messages.tool_calls.index of type int","type":"invalid_request_error","param":null,"code":null}}]”}],“type”:“illegal_argument_exception”,“reason”:“Agentic search failed - Agent execution error - Agent ID: [_Nh7rZsBMCptIK-aGFFT], Error: [Error from remote service: {"error":{"message":"j**son: cannot unmarshal number 0.0 into Go struct field ToolCall.messages.tool_calls.index of type int",**"type":"invalid_request_error","param":null,"code":null}}]”,“caused_by”:{“type”:“status_exception”,“reason”:“Error from remote service: {"error":{"message":"json: cannot unmarshal number 0.0 into Go struct field ToolCall.messages.tool_calls.index of type int","type":"invalid_request_error","param":null,"code":null}}”}},“status”:400
Configuration:
OS/Hardware: MacOS: MacBook Pro M3 Max OpenSearch: OpenSearch 3.4 running under Docker. Image: opensearchproject/opensearch:3.4.0 LLM: A Qwen model running under Ollama. Ollama is running on host. Ollam version: 0.13.5
** Plugins **
opensearch-node1 opensearch-alerting 3.4.0.0 opensearch-node1 opensearch-anomaly-detection 3.4.0.0 opensearch-node1 opensearch-asynchronous-search 3.4.0.0 opensearch-node1 opensearch-cross-cluster-replication 3.4.0.0 opensearch-node1 opensearch-custom-codecs 3.4.0.0 opensearch-node1 opensearch-flow-framework 3.4.0.0 opensearch-node1 opensearch-geospatial 3.4.0.0 opensearch-node1 opensearch-index-management 3.4.0.0 opensearch-node1 opensearch-job-scheduler 3.4.0.0 opensearch-node1 opensearch-knn 3.4.0.0 opensearch-node1 opensearch-ltr 3.4.0.0 opensearch-node1 opensearch-ml 3.4.0.0 opensearch-node1 opensearch-neural-search 3.4.0.0 opensearch-node1 opensearch-notifications 3.4.0.0 opensearch-node1 opensearch-notifications-core 3.4.0.0 opensearch-node1 opensearch-observability 3.4.0.0 opensearch-node1 opensearch-performance-analyzer 3.4.0.0 opensearch-node1 opensearch-reports-scheduler 3.4.0.0 opensearch-node1 opensearch-search-relevance 3.4.0.0 opensearch-node1 opensearch-security 3.4.0.0 opensearch-node1 opensearch-security-analytics 3.4.0.0 opensearch-node1 opensearch-skills 3.4.0.0 opensearch-node1 opensearch-sql 3.4.0.0 opensearch-node1 opensearch-system-templates 3.4.0.0 opensearch-node1 opensearch-ubi 3.4.0.0 opensearch-node1 query-insights 3.4.0.0
HTTP Traffic:
I sniffed the HTTP traffic between OpenSearch and Ollama and found that OpenSearch is indeed sending an index property with a float (0.0) instead of an integer (0).
OpenSearch sends 2 POSTs to the v1/chat/completions endpoint. The response to the first POST looks like the below (note the index property in choices[0].tool_calls[0].index):
{
"id": "chatcmpl-170",
"object": "chat.completion",
"created": 1768222804,
"model": "qwen3-a3b-16k",
"system_fingerprint": "fp_ollama",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "",
"tool_calls": [
{
"id": "call_2l0j5wr2",
"index": 0, <---- Note!
"type": "function",
"function": {
"name": "ListIndexTool",
"arguments": "{\"indices\":[\"able_chunks_v1\"]}"
}
}
]
},
"finish_reason": "tool_calls"
}
],
"usage": {
"prompt_tokens": 1643,
"completion_tokens": 25,
"total_tokens": 1668
}
}
The second POST contains the following snippet (I've elided the rest of the request to save space, but can provide it if needed). Note that index property under tool_calls[0]:
{
"role": "assistant",
"content": "",
"tool_calls": [
{
"id": "call_2l0j5wr2",
"index": 0.0, <---- Note!
"type": "function",
"function": {
"name": "ListIndexTool",
"arguments": "{\"indices\":[\"able_chunks_v1\"]}"
}
}
]
},
So it looks like OpenSearch is formatting that value as a float instead of an integer which is causing the problem
Additional information. Some Googling came up with the statement below. So perhaps a fix for this is not to include the index property at all.
In the OpenAI API protocol, the index property is primarily a feature of streaming responses (via the ChatCompletionChunk object). In a standard non-streaming request, the assistant message contains a tool_calls array where order is implicit by the list index. However, when you send a message back to the API to maintain conversation history, or if you are manually constructing a request that reflects a previous tool-calling turn, you do not typically include the index property in the messages array; the API expects id, type, and function
@jiapingzeng can you confirm your json fix will fix this issue on 3.4 version?
Was able to reproduce this issue on 3.4 with Ollama running Qwen3:30b.
Tried on 3.5 and I am getting a different error 😅
{
"error": {
"root_cause": [
{
"type": "illegal_argument_exception",
"reason": "Agentic search failed - Agent execution error - Agent ID: [XLhVDpwBnKaP40sIXPja], Error: [No valid 'dsl_query' found in conversational agent response. The agent must return a JSON object with 'dsl_query' field. Please check the agent configuration and prompts.]"
}
],
"type": "illegal_argument_exception",
"reason": "Agentic search failed - Agent execution error - Agent ID: [XLhVDpwBnKaP40sIXPja], Error: [No valid 'dsl_query' found in conversational agent response. The agent must return a JSON object with 'dsl_query' field. Please check the agent configuration and prompts.]",
"caused_by": {
"type": "illegal_argument_exception",
"reason": "No valid 'dsl_query' found in conversational agent response. The agent must return a JSON object with 'dsl_query' field. Please check the agent configuration and prompts."
}
},
"status": 400
}
But at least the serialization issue seems to be fixed by https://github.com/opensearch-project/ml-commons/pull/4520 based on tcpdump:
{
"id": "chatcmpl-454",
"object": "chat.completion",
"created": 1769769821,
"model": "qwen3:30b",
"system_fingerprint": "fp_ollama",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "",
"reasoning": "...",
"tool_calls": [
{
"id": "call_wna7f3fo",
"index": 0, <---- No longer a float
"type": "function",
"function": {
"name": "QueryPlanningTool",
"arguments": "{\"index_name\":\"iris-index\",\"question\":\"List all the flowers present\"}"
}
}
]
},
"finish_reason": "tool_calls"
}
],
"usage": {
"prompt_tokens": 377,
"completion_tokens": 380,
"total_tokens": 757
}
}
We will backport this PR to 3.4 to fix there too.
