[GH-ISSUE #24076] issue: RAG template: duplicate {{CONTEXT}} / [context] placeholders inject retrieved context multiple times without warning #58843

New Issue

GiteaMirror · 2026-05-06T00:16:34-05:00

GiteaMirror commented

2026-05-06 00:16:34 -05:00

Originally created by @KingsleyOWO on GitHub (Apr 24, 2026).
Original GitHub issue: https://github.com/open-webui/open-webui/issues/24076

Check Existing Issues

I have searched for any existing and/or related issues.
I have searched for any existing and/or related discussions.
I have also searched in the CLOSED issues AND CLOSED discussions and found no related items (your issue might already be addressed on the development branch!).
I am using the latest version of Open WebUI.

Installation Method

Docker

Open WebUI Version

v0.9.0

Ollama Version (if applicable)

No response

Operating System

Ubuntu 24.04

Browser (if applicable)

No response

Confirmation

I have read and followed all instructions in README.md.
I am using the latest version of both Open WebUI and Ollama.
I have included the browser console logs.
I have included the Docker container logs.
I have provided every relevant configuration, setting, and environment variable used in my setup.
I have clearly listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc).
I have documented step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation. My steps:
Start with the initial platform/version/OS and dependencies used,
Specify exact install/launch/configure commands,
List URLs visited, user input (incl. example values/emails/passwords if needed),
Describe all options and toggles enabled or changed,
Include any files or environmental changes,
Identify the expected and actual result at each stage,
Ensure any reasonably skilled user can follow and hit the same issue.

Expected Behavior

When a RAG template contains more than one context placeholder, OpenWebUI should warn that retrieved context will be injected once per placeholder.

Acceptable fixes could include either:

a backend warning from rag_template() when the rendered template has context_count > 1, or
a UI-side warning/lint when saving a custom RAG template with duplicate [context] / {{CONTEXT}} placeholders.

The warning should not block saving and should not change existing behavior, because some users may intentionally duplicate context placeholders.

Actual Behavior

A template with duplicate context placeholders injects the full retrieved context multiple times silently.

There is already a debug warning when no context placeholder exists:

if '[context]' not in template and '{{CONTEXT}}' not in template:
    log.debug("WARNING: The RAG template does not contain the '[context]' or '{{CONTEXT}}' placeholder.")

However, there is no corresponding warning when more than one context placeholder exists. The missing-placeholder case and duplicate-placeholder case are both easy-to-miss RAG-template failure modes, but only the missing-placeholder case currently emits a signal.

Steps to Reproduce

Go to Admin Settings → Documents → RAG Template.
Save a custom RAG template that contains {{CONTEXT}} twice:

### Task
Answer using the provided context.

<context>
{{CONTEXT}}
</context>

Later in the same template:

```xml
<context>
{{CONTEXT}}
</context>
```

Run any RAG-enabled chat turn.
Capture the rendered prompt reaching the model, for example by using a local proxy, enabling debug instrumentation, or temporarily logging the output of rag_template().
Observe that the retrieved context is inserted once for each placeholder occurrence.

Minimal substitution-only reproducer:

template = "A {{CONTEXT}} B {{CONTEXT}}"
context = "CTX"

rendered = template.replace("{{CONTEXT}}", context)

assert rendered == "A CTX B CTX"

The OpenWebUI code path does the same kind of global replacement for both [context] and {{CONTEXT}}.

Logs & Screenshots

No browser-console error is expected for this issue. This is a rendered-prompt/template-expansion behavior, not a frontend crash.

Browser console:

No relevant frontend error observed.

Container / backend logs:

No exception observed when the duplicate-placeholder RAG template is used.
The issue is that there is also no warning/log when the template contains more than one context placeholder.
The duplicate injection is only visible if the final rendered prompt is inspected or token-counted.

Minimal rendered-output evidence:

Template:

A
{{CONTEXT}}
B
{{CONTEXT}}
C

Context:

<source id="1">example retrieved chunk</source>

Rendered result:

A
<source id="1">example retrieved chunk</source>
B
<source id="1">example retrieved chunk</source>
C

This is consistent with the current rag_template() substitution path, which uses:

template = template.replace('[context]', context)
template = template.replace('{{CONTEXT}}', context)

without a count argument.

Token-accounting reproduction from my environment:

Metric	Template with 1× `{{CONTEXT}}`	Template with 2× `{{CONTEXT}}`	Delta	One context copy
chars	4,985	8,687	+3,702	3,551
Qwen tokenizer	3,024	5,843	+2,819	2,710
tiktoken `o200k_base`	3,450	6,701	+3,251	3,141

The rendered prompt therefore contains one extra full copy of the retrieved context. The current UI/logs do not make this visible.

Additional Information

Related but not the same as existing RAG duplication reports:

#19098 discusses prompt/context duplication when a RAG template is used with tool results.
#21167 discusses user-message duplication for RAG queries.
#21726 discusses RAG duplication when view_knowledge_file is used.

This issue is narrower: a custom RAG template that itself contains duplicate {{CONTEXT}} / [context] placeholders causes the retrieved context to be injected once per placeholder.

I am not asking to change substitution semantics. Intentional duplicate placeholders should continue to work. I am only requesting warning/linting so that accidental duplication is visible.

Anecdotal observation, not part of the bug claim:

In one internal Traditional Chinese HR/policy RAG deployment, the duplicate-context template appeared to produce better answers than the default single-context template in a small manual comparison. In particular, it handled multi-clause policy distinctions more reliably.

However, I am not claiming this proves duplicate injection is generally beneficial. The comparison was not controlled enough: the custom template also differed in language, structure, citation guidance, and XML/code-fence formatting; the sample size was only n=2; and only one model family was tested.

This observation is included only to explain why a user might unknowingly keep a duplicate-context template in production: it may appear to improve answer quality while silently increasing prompt tokens. A warning would let users make that tradeoff intentionally.

Originally created by @KingsleyOWO on GitHub (Apr 24, 2026). Original GitHub issue: https://github.com/open-webui/open-webui/issues/24076 ### Check Existing Issues - [x] I have searched for any existing and/or related issues. - [x] I have searched for any existing and/or related discussions. - [x] I have also searched in the CLOSED issues AND CLOSED discussions and found no related items (your issue might already be addressed on the development branch!). - [x] I am using the latest version of Open WebUI. ### Installation Method Docker ### Open WebUI Version v0.9.0 ### Ollama Version (if applicable) _No response_ ### Operating System Ubuntu 24.04 ### Browser (if applicable) _No response_ ### Confirmation - [x] I have read and followed all instructions in `README.md`. - [x] I am using the latest version of **both** Open WebUI and Ollama. - [x] I have included the browser console logs. - [x] I have included the Docker container logs. - [x] I have **provided every relevant configuration, setting, and environment variable used in my setup.** - [x] I have clearly **listed every relevant configuration, custom setting, environment variable, and command-line option that influences my setup** (such as Docker Compose overrides, .env values, browser settings, authentication configurations, etc). - [x] I have documented **step-by-step reproduction instructions that are precise, sequential, and leave nothing to interpretation**. My steps: - Start with the initial platform/version/OS and dependencies used, - Specify exact install/launch/configure commands, - List URLs visited, user input (incl. example values/emails/passwords if needed), - Describe all options and toggles enabled or changed, - Include any files or environmental changes, - Identify the expected and actual result at each stage, - Ensure any reasonably skilled user can follow and hit the same issue. ### Expected Behavior When a RAG template contains more than one context placeholder, OpenWebUI should warn that retrieved context will be injected once per placeholder. Acceptable fixes could include either: 1. a backend warning from `rag_template()` when the rendered template has `context_count > 1`, or 2. a UI-side warning/lint when saving a custom RAG template with duplicate `[context]` / `{{CONTEXT}}` placeholders. The warning should not block saving and should not change existing behavior, because some users may intentionally duplicate context placeholders. ### Actual Behavior A template with duplicate context placeholders injects the full retrieved context multiple times silently. There is already a debug warning when no context placeholder exists: ```python if '[context]' not in template and '{{CONTEXT}}' not in template: log.debug("WARNING: The RAG template does not contain the '[context]' or '{{CONTEXT}}' placeholder.") ``` However, there is no corresponding warning when more than one context placeholder exists. The missing-placeholder case and duplicate-placeholder case are both easy-to-miss RAG-template failure modes, but only the missing-placeholder case currently emits a signal. ### Steps to Reproduce 1. Go to Admin Settings → Documents → RAG Template. 2. Save a custom RAG template that contains `{{CONTEXT}}` twice: ````text ### Task Answer using the provided context. <context> {{CONTEXT}} </context> Later in the same template: ```xml <context> {{CONTEXT}} </context> ``` ```` 3. Run any RAG-enabled chat turn. 4. Capture the rendered prompt reaching the model, for example by using a local proxy, enabling debug instrumentation, or temporarily logging the output of `rag_template()`. 5. Observe that the retrieved context is inserted once for each placeholder occurrence. Minimal substitution-only reproducer: ```python template = "A {{CONTEXT}} B {{CONTEXT}}" context = "CTX" rendered = template.replace("{{CONTEXT}}", context) assert rendered == "A CTX B CTX" ``` The OpenWebUI code path does the same kind of global replacement for both `[context]` and `{{CONTEXT}}`. ### Logs & Screenshots No browser-console error is expected for this issue. This is a rendered-prompt/template-expansion behavior, not a frontend crash. Browser console: - No relevant frontend error observed. Container / backend logs: - No exception observed when the duplicate-placeholder RAG template is used. - The issue is that there is also no warning/log when the template contains more than one context placeholder. - The duplicate injection is only visible if the final rendered prompt is inspected or token-counted. Minimal rendered-output evidence: Template: ```text A {{CONTEXT}} B {{CONTEXT}} C ``` Context: ```text <source id="1">example retrieved chunk</source> ``` Rendered result: ```text A <source id="1">example retrieved chunk</source> B <source id="1">example retrieved chunk</source> C ``` This is consistent with the current `rag_template()` substitution path, which uses: ```python template = template.replace('[context]', context) template = template.replace('{{CONTEXT}}', context) ``` without a `count` argument. Token-accounting reproduction from my environment: | Metric | Template with 1× `{{CONTEXT}}` | Template with 2× `{{CONTEXT}}` | Delta | One context copy | | --- | ---: | ---: | ---: | ---: | | chars | 4,985 | 8,687 | +3,702 | 3,551 | | Qwen tokenizer | 3,024 | 5,843 | +2,819 | 2,710 | | tiktoken `o200k_base` | 3,450 | 6,701 | +3,251 | 3,141 | The rendered prompt therefore contains one extra full copy of the retrieved context. The current UI/logs do not make this visible. ### Additional Information Related but not the same as existing RAG duplication reports: - #19098 discusses prompt/context duplication when a RAG template is used with tool results. - #21167 discusses user-message duplication for RAG queries. - #21726 discusses RAG duplication when `view_knowledge_file` is used. This issue is narrower: a custom RAG template that itself contains duplicate `{{CONTEXT}}` / `[context]` placeholders causes the retrieved context to be injected once per placeholder. I am not asking to change substitution semantics. Intentional duplicate placeholders should continue to work. I am only requesting warning/linting so that accidental duplication is visible. Anecdotal observation, not part of the bug claim: In one internal Traditional Chinese HR/policy RAG deployment, the duplicate-context template appeared to produce better answers than the default single-context template in a small manual comparison. In particular, it handled multi-clause policy distinctions more reliably. However, I am not claiming this proves duplicate injection is generally beneficial. The comparison was not controlled enough: the custom template also differed in language, structure, citation guidance, and XML/code-fence formatting; the sample size was only n=2; and only one model family was tested. This observation is included only to explain why a user might unknowingly keep a duplicate-context template in production: it may appear to improve answer quality while silently increasing prompt tokens. A warning would let users make that tradeoff intentionally.

GiteaMirror added the bug label 2026-05-06 00:16:34 -05:00

GiteaMirror closed this issue

2026-05-06 00:16:36 -05:00

GiteaMirror commented

2026-05-06 00:16:37 -05:00

@tjbck commented on GitHub (Apr 24, 2026):

Intended behaviour, duplication notice message added to dev.

@tjbck commented on GitHub (Apr 24, 2026): Intended behaviour, duplication notice message added to dev.

Sign in to join this conversation.

Branches Tags

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: github-starred/open-webui#58843

[GH-ISSUE #24076] issue: RAG template: duplicate {{CONTEXT}} / [context] placeholders inject retrieved context multiple times without warning #58843