From a05b3fbd335fdd89527d9dbcbf79f356fa604bd7 Mon Sep 17 00:00:00 2001
From: vegu-ai-tools <152010387+vegu-ai-tools@users.noreply.github.com>
Date: Fri, 1 May 2026 13:52:17 +0300
Subject: [PATCH] fix: correct volatile context placement in RAG query
 generation template

---
 CHANGELOG.yaml                                |  1 +
 ...alyze-text-and-generate-rag-queries.jinja2 | 22 +++++++++++++++----
 2 files changed, 19 insertions(+), 4 deletions(-)

diff --git a/CHANGELOG.yaml b/CHANGELOG.yaml
index c59a2115..60552616 100644
--- a/CHANGELOG.yaml
+++ b/CHANGELOG.yaml
@@ -49,6 +49,7 @@
   - "Game Loop Actor Iter: Fixed a bug where the `game_loop_actor_iter` event was only firing for player turns, so agents that listen for actor iteration (like passive narration) silently skipped AI turns."
   - "OpenRouter: Fixed redundant provider/model list fetches when multiple OpenRouter clients are configured. Concurrent `status()` calls all saw the fetched-flag as still false and each kicked off its own HTTP request; the fetches are now serialized with an asyncio lock and double-checked locking so only one request runs."
   - "Character Prompts: Fixed duplicated character description in prompts. When the LLM-generated character sheet contained a `description` attribute (common — the extractor often emits one), it was rendered alongside the standalone description field, repeating the same text twice. Templates that render both now filter `description` out of the sheet (case-insensitive). Affects the dialogue, narrator, director (voice assignment), and creator (dialogue examples, dialogue instructions, character goals, contextual generate) prompt paths."
+  - "RAG Query Generation: Fixed the analyze-text-and-generate-rag-queries template (used by the long-term memory retrieval flow) ignoring the volatile context placement setting. Dynamic instructions and the passed-in RAG context were always pinned at the top of the prompt, defeating prompt caching even when it was enabled on the client/agent. Volatile pieces are now split out and positioned via volatile_context_placement, matching the rest of the prompt set."
   improvements:
   - "System Prompt Override Indicators: The system prompt override list now shows a pencil icon next to entries that have an active override, making it easy to see which prompts have been customized."
   - "Search Strictness: The distance_mod embedding preset value is now a float (was int) with a range of 0.1–2.0, allowing both tighter and looser similarity matching. A 'Search Strictness' slider is now available in the Context Database UI, letting users tune search sensitivity on the fly. Changes persist to the active embedding preset immediately."
diff --git a/src/talemate/prompts/templates/world_state/analyze-text-and-generate-rag-queries.jinja2 b/src/talemate/prompts/templates/world_state/analyze-text-and-generate-rag-queries.jinja2
index 9405cccd..60e9a7b3 100644
--- a/src/talemate/prompts/templates/world_state/analyze-text-and-generate-rag-queries.jinja2
+++ b/src/talemate/prompts/templates/world_state/analyze-text-and-generate-rag-queries.jinja2
@@ -10,20 +10,34 @@ Override example:
 {% set rendered_context -%}
 <|SECTION:CONTEXT|>
 {% include "extra-context-static.jinja2" %}
-{% include "extra-context-dynamic.jinja2" %}
-{% if extra_context %}{% for context in extra_context %}{{ context }}
----
-{% endfor %}{% endif %}
 <|CLOSE_SECTION|>
 {% if include_character_context %}{% include "character-context.jinja2" %}{% endif %}
 {% endset %}
 {# END RENDERED CONTEXT #}
 
+{# VOLATILE CONTEXT #}
+{% set volatile_inner -%}
+{% include "extra-context-dynamic.jinja2" %}
+{% if extra_context %}{% for context in extra_context %}{{ context }}
+---
+{% endfor %}{% endif %}
+{%- endset %}
+{% set volatile_context_text -%}
+{% if volatile_inner.strip() %}<|SECTION:DYNAMIC CONTEXT|>
+{{ volatile_inner }}
+<|CLOSE_SECTION|>
+{% endif %}
+{%- endset %}
+{# END VOLATILE CONTEXT #}
+
 {# RENDER PROMPT #}
+{% set volatile_placement = volatile_context_placement() %}
 {{ rendered_context }}
+{% if volatile_placement != "after_history" %}{{ volatile_context_text }}{% endif %}
 <|SECTION:SCENE|>
 {{ text }}
 <|CLOSE_SECTION|>
+{% if volatile_placement == "after_history" %}{{ volatile_context_text }}{% endif %}
 <|SECTION:TASK|>
 You are assisting with an ongoing story. You have access to a vector database containing factual information about the characters, locations, events, and lore of this narrative world. Your task is to generate up to {{ num_queries }} specific, targeted queries to gather additional context for the current scene or conversation.