Made refiner, referencer and builder less strict

prompts/video_references.md

@@ -39,6 +39,8 @@
 
 `ids` должны содержать только ID событий из `present`, во время которых нужно
 искать содержимое. ID указывай в хронологическом порядке и без повторов.
+Если нужный фрагмент пересекает границу окна, укажи только его ID из `present`;
+если таких ID нет, не создавай ссылку.
 
 `text` — короткая конкретная инструкция следующему этапу:
 

structure_builder.py

@@ -1,4 +1,5 @@
 import json
+import logging
 import traceback
 from typing import Annotated, Literal
 
@@ -180,6 +181,49 @@ class StructureBuilder:
                 "pending.event_ids must reference events from the current window"
             )
 
+    @staticmethod
+    def _keep_window_references(
+        result: _BuildResult,
+        window: Window[Event],
+    ) -> _BuildResult:
+        """Drop impossible image references and repair technical pending IDs."""
+        visual_ids = {
+            event.id
+            for event in window.past + window.present
+            if event.type == "vis"
+        }
+        elements: list[StructureElement] = []
+        for element in result.elements:
+            if isinstance(element, ImageElement) and element.event_id not in visual_ids:
+                logging.warning("Ignoring image outside the current window: id=%s", element.event_id)
+                continue
+            elements.append(element)
+
+        context = result.context
+        if context.pending is not None:
+            window_ids = {
+                event.id
+                for event in window.past + window.present + window.future
+            }
+            pending_ids = [
+                event_id
+                for event_id in context.pending.event_ids
+                if event_id in window_ids
+            ]
+            if pending_ids != context.pending.event_ids:
+                logging.warning(
+                    "Trimming pending IDs to the current window: %s -> %s",
+                    context.pending.event_ids,
+                    pending_ids,
+                )
+                pending = (
+                    context.pending.model_copy(update={"event_ids": pending_ids})
+                    if pending_ids else None
+                )
+                context = context.model_copy(update={"pending": pending})
+
+        return result.model_copy(update={"elements": elements, "context": context})
+
     def _build_window(self, window: Window[Event]) -> _BuildResult:
         """Build document elements from a single timeline window."""
         messages = [
@@ -211,6 +255,7 @@ class StructureBuilder:
 
             try:
                 result = _BuildResult.model_validate_json(response)
+                result = self._keep_window_references(result, window)
                 self._validate_result(result, window)
                 return result
             except (ValidationError, ValueError, TypeError):

structure_refiner.py

@@ -1,4 +1,5 @@
 import json
+import logging
 import traceback
 
 from pydantic import BaseModel, ConfigDict, Field, ValidationError, model_validator
@@ -58,6 +59,25 @@ class StructureRefiner:
             if isinstance(element, ImageElement) and element.event_id not in input_image_ids:
                 raise ValueError("image.event_id must come from present")
 
+    @staticmethod
+    def _keep_present_images(
+        result: _RefineResult,
+        window: Window[StructureElement],
+    ) -> _RefineResult:
+        """Discard image elements invented from context windows."""
+        input_image_ids = {
+            element.event_id
+            for element in window.present
+            if isinstance(element, ImageElement)
+        }
+        elements: list[StructureElement] = []
+        for element in result.elements:
+            if isinstance(element, ImageElement) and element.event_id not in input_image_ids:
+                logging.warning("Ignoring image outside present window: id=%s", element.event_id)
+                continue
+            elements.append(element)
+        return result.model_copy(update={"elements": elements})
+
     def _refine_window(self, window: Window[StructureElement]) -> _RefineResult:
         messages = [
             self._system_prompt,
@@ -91,6 +111,7 @@ class StructureRefiner:
 
             try:
                 result = _RefineResult.model_validate_json(response)
+                result = self._keep_present_images(result, window)
                 self._validate_result(result, window)
                 return result
             except (ValidationError, ValueError, TypeError):

video_references.py

@@ -1,4 +1,5 @@
 import json
+import logging
 import traceback
 from typing import Literal
 
@@ -99,6 +100,47 @@ class VideoReferenceBuilder:
                 raise ValueError("Duplicate reference in a single window")
             seen.add(key)
 
+    @staticmethod
+    def _keep_present_references(
+        result: _WindowResult,
+        window: Window[VideoReferenceEvent],
+    ) -> _WindowResult:
+        """Keep the usable part of references that cross a window boundary."""
+        present_positions = {
+            event.id: position
+            for position, event in enumerate(window.present)
+        }
+        references: list[UnresolvedReference] = []
+        seen: set[tuple[tuple[int, ...], str]] = set()
+
+        for reference in result.unresolved:
+            ids = sorted(
+                {event_id for event_id in reference.ids if event_id in present_positions},
+                key=present_positions.__getitem__,
+            )
+            if not ids:
+                logging.warning(
+                    "Ignoring reference outside present window: ids=%s",
+                    reference.ids,
+                )
+                continue
+            if ids != reference.ids:
+                logging.warning(
+                    "Trimming reference to present window: ids=%s -> %s",
+                    reference.ids,
+                    ids,
+                )
+
+            normalized = reference.model_copy(update={"ids": ids})
+            key = (tuple(ids), normalized.type)
+            if key in seen:
+                logging.warning("Ignoring duplicate reference: ids=%s", ids)
+                continue
+            seen.add(key)
+            references.append(normalized)
+
+        return result.model_copy(update={"unresolved": references})
+
     def _build_window(
         self,
         window: Window[VideoReferenceEvent],
@@ -135,6 +177,7 @@ class VideoReferenceBuilder:
 
             try:
                 result = _WindowResult.model_validate_json(response)
+                result = self._keep_present_references(result, window)
                 self._validate_result(result, window)
                 return result
             except (ValidationError, ValueError, TypeError):