web search not working or not intuitive

[im360john]

What happened?

installed serper, firecrawl and cohere api keys into env variables.

when trying the web search with gpt 4 mini or claude 3.7, i get a cannot search the web response:

also, does agents have access to the same web search or is it limited to google's api path as a tool

Version Information

latest

Steps to Reproduce

see above

What browsers are you seeing the problem on?

No response

Relevant log output

na

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

try to toggle the "Search" button on/off, and check server logs

[im360john]

nothing in the logs. tried toggling on and off. also tried incognito mode

[doprado]

I am facing the same problem.

[im360john]

{
baseURL: "https://api.openai.com/v1",
modelOptions.model: "gpt-4o-mini",
modelOptions.user: "6834a2822f462c3d2a1b3b77",
modelOptions.stream: true,
// 1 message(s)
modelOptions.messages: [{"role":"user","content":"google the best prices"}],
}
2025-05-26T21:36:30.078Z debug: Making request to https://api.openai.com/v1/chat/completions
2025-05-26T21:36:30.083Z debug: [saveConvo] api/app/clients/BaseClient.js - saveMessageToDatabase #saveConvo
2025-05-26T21:36:30.117Z debug: [FinalizationRegistry] Cleaning up client for user 68267c068706445bb3167dc9
2025-05-26T21:36:30.117Z debug: [FinalizationRegistry] Cleaning up client for user 6834a2822f462c3d2a1b3b77
2025-05-26T21:36:33.613Z debug: [OpenAIClient] chatCompletion response
{
object: "chat.completion",
service_tier: "default",
usage.prompt_tokens: 11,
usage.completion_tokens: 231,
usage.total_tokens: 242,
prompt_tokens_details.cached_tokens: 0,
prompt_tokens_details.audio_tokens: 0,
completion_tokens_details.reasoning_tokens: 0,
completion_tokens_details.audio_tokens: 0,
completion_tokens_details.accepted_prediction_tokens: 0,
completion_tokens_details.rejected_prediction_tokens: 0,
id: "chatcmpl-BbZmgyEr34F7m54z9stwhuCCbUihI",
// 1 choice(s)
choices: [{"message":{"role":"assistant","content":"I can't browse the internet in real time, but I can sugges... [truncated]],
created: 1748295390,
model: "gpt-4o-mini-2024-07-18",
system_fingerprint: "fp_34a54ae93c",
}
2025-05-26T21:36:33.622Z debug: [spendTokens] conversationId: 3ceca4fe-7070-4b02-bec2-7f4a365d3133 | Context: message | Token usage:
{
promptTokens: 11,
completionTokens: 231,
}
2025-05-26T21:36:33.660Z debug: [spendTokens] Transaction data record against balance:
{
user: "6834a2822f462c3d2a1b3b77",
prompt: -1.65,
promptRate: 0.15,
completion: -138.6,
completionRate: 0.6,
balance: 199829.9,
}
2025-05-26T21:36:33.660Z debug: [spendTokens] conversationId: 3ceca4fe-7070-4b02-bec2-7f4a365d3133 | Context: reasoning | Token usage:
{
promptTokens: undefined,
completionTokens: 0,
}
2025-05-26T21:36:33.673Z debug: [spendTokens] Transaction data record against balance:
{
user: "6834a2822f462c3d2a1b3b77",
prompt: undefined,
promptRate: undefined,
completion: 0,
completionRate: 0.6,
balance: 199829.9,
}
2025-05-26T21:36:33.681Z debug: [saveConvo] api/app/clients/BaseClient.js - saveMessageToDatabase #saveConvo
2025-05-26T21:36:33.693Z debug: [AskController] Request closed
2025-05-26T21:36:33.701Z debug: [AskController] Performing cleanup
2025-05-26T21:36:33.701Z debug: [AskController] Cleaning up abort controller
2025-05-26T21:36:33.701Z debug: [AskController] Cleanup completed

--

nothing specific about the search tool being triggered or any of the apis (serper, cohere, firecrawl) being called.

[danny-avila]

I think I know the issue, what is the value of your ENDPOINTS environment variable? For now, agents needs to be included, but I will think about this case where it may not be set.

ENDPOINTS=openAI,agents

[im360john]

libreyaml.txt
hmm, agents on claude or openai still responding that they can't search the web. Anything wrong from my attached librechat.yaml file? My env variables has ENDPOINTS set to custom,agents,openAI and I have the keys for serper, etc.

nothing in the log that indicates errors

[doprado]

@im360john
I believe something is missing.

allowedProviders:
 - openAI

[im360john]

tried adding and still the same response.
allowedProviders:
- openAI
- anthropic
- google

[danny-avila]

@im360john the config looks good and should work. To be sure webSearch wasn't disabled (although you shouldn't see it if it were), you can add webSearch: true to interface:

interface:
  runCode: true
  agents: true
  webSearch: true
  modelSelect: true
  customWelcome: "Hey {{user.name}}! Welcome"

Also, what does your ENDPOINTS environment variable look like? For now, it should include agents, but I'm going to handle the case where agents endpoint is not defined.

# example
ENDPOINTS=openAI,agents
# instead of:
# ENDPOINTS=openAI

See changes merging soon: #7598

I'm also adding better logging for when the capability is disabled (and therefore not being added to the chat run).

[Odrec]

web_search is a capability that needs to be added:

endpoints:
  agents:
    recursionLimit: 50
    maxRecursionLimit: 100
    disableBuilder: false
    capabilities:
      - "file_search"
      - "artifacts"
      - "actions"
      - "tools"
      - "web_search"

https://www.librechat.ai/docs/features/web_search#usage https://www.librechat.ai/docs/configuration/librechat_yaml/object_structure/agents#capabilities

it worked!

[im360john]

I tried adding ENDPOINTS=custom,agents,openAI and my agents still responds with no capability of web search. here is the agents portion of my yaml file:

endpoints:
agents:
recursionLimit: 50
maxRecursionLimit: 100
disableBuilder: false
capabilities:
- "actions"
- "execute_code"
- "artifacts"
- "tools"
- "chain"
- "file_search"
- "ocr"

[Odrec]

You need to add web_search to the capabilities

#7581 (reply in thread)

[rubenaster]

Is there a more lightweight alternative?

Especially in enterprise environment you can just register and use four new 3rd party services. And the web search feature is so fundamentally important.

[alexanderwbg]

https://github.com/ItzCrazyKns/Perplexica is an open-source alternative to perplexity, runs on 2.3gb VRAM

[rubenaster]

Oh wow, definitely will check that one out! Thanks a bunch.

[danny-avila]

@rubenaster perplexica is not more lightweight as you are adding another full stack project to the stack. I think what you are asking for is "open source." We are adding open source alternatives for each of the search "components" but running all the necessary parts in an open-source manner is not always lightweight.

[danny-avila]

The docs specify "planned" providers for each of the components:
https://www.librechat.ai/docs/features/web_search#1-search-providers

[guygud]

Excuse me, but I think I’m not understanding something. When I turn on web search, the interface suggests adding API keys for SERPER, Firecrawl, and Jina. However, that’s not exactly what I need. I just want to have the links from the citations block (like in Perplexity) displayed in the interface. Is there any way to enable this?

[danny-avila]

This feature is a native search feature that has nothing to do with Perplexity.

https://www.librechat.ai/docs/features/web_search

I know there are some people interested in using the same citation UI I created for use when using Perplexity models, but you will have to defer to them if/when they contribute this.

[guygud]

Oh, I see. It's just that in the thread dedicated to Perplexity, there was a link to this feature.

And it seemed like someone tried to submit a feature for Perplexity support, but it was rejected.

Okay, now I understand that I need to wait for something similar to be implemented for Perplexity and gpt-web-search.

Could you please share a link to the relevant feature request, so there’s no confusion? Thanks!

[LCBO]

Regarding the search, I agree that is fundamental. So far, from all projects I tried, the most advanced using the search tools is this one:
https://github.com/khoj-ai/khoj

They have also a demo.

The most impressive is the fact that the system decides when it needs more information to be searched, define the perfect search strings for the search engine, read the pages using even internal scrapes, see if it has enough information and provide answer. If not, it will search more. It is very fast.

The bad point for the project is the difficult manual installation (used for dev). The docker is fine though.

[CorneeldH]

I tried all of the above with local searXNG, but can't get it to work..

Any pointers?

.env:
env_copy.txt

librechat.yaml:
librechat_copy.yaml.txt

[CorneeldH]

When I add a serper api key within the agent builder it does work.. But not with OpenAI gpt-4o-mini and the api key added

[CorneeldH]

It also works with SearXNG via the agents. Just not via the chat. I will try a few more chat models next to OpenAI.

[danny-avila]

Since I didn't see you mention it, have you looked at the documentation?

https://www.librechat.ai/docs/configuration/librechat_yaml/object_structure/web_search#setting-up-searxng

[CorneeldH]

It works now! The thing that threw me off was that in chat for some reason I got the settings when I clicked on the search button (instead of activating it). At least, that's how it felt.. Now I can't reproduce of course..

I played with it today, including setting up mcp's etc, looks really good!