From 9a78882b5a294fa7de824197d08a743810a6e777 Mon Sep 17 00:00:00 2001
From: Trenton H <797416+stumpylog@users.noreply.github.com>
Date: Wed, 13 May 2026 09:01:34 -0700
Subject: [PATCH] Fix: Don't embed the metadata which is already embedded into
 the context (#12795)

---
 src/paperless_ai/indexing.py               | 10 +++++++++-
 src/paperless_ai/tests/test_ai_indexing.py | 20 ++++++++++++++++++++
 2 files changed, 29 insertions(+), 1 deletion(-)

diff --git a/src/paperless_ai/indexing.py b/src/paperless_ai/indexing.py
index 9293633a3..d596805fa 100644
--- a/src/paperless_ai/indexing.py
+++ b/src/paperless_ai/indexing.py
@@ -116,7 +116,15 @@ def build_document_node(document: Document) -> list["BaseNode"]:
     from llama_index.core import Document as LlamaDocument
     from llama_index.core.node_parser import SimpleNodeParser
 
-    doc = LlamaDocument(text=text, metadata=metadata)
+    # Exclude all metadata keys from the embedding text — build_llm_index_text
+    # already encodes this info in the body, so prepending it again would double
+    # the token count and exceed embedding models with small context windows
+    # (e.g. nomic-embed-text via Ollama defaults to num_ctx=2048).
+    doc = LlamaDocument(
+        text=text,
+        metadata=metadata,
+        excluded_embed_metadata_keys=list(metadata.keys()),
+    )
     parser = SimpleNodeParser(
         chunk_size=RAG_CHUNK_SIZE,
         chunk_overlap=get_rag_chunk_overlap(),
diff --git a/src/paperless_ai/tests/test_ai_indexing.py b/src/paperless_ai/tests/test_ai_indexing.py
index 09fbc2038..aba9e2f7f 100644
--- a/src/paperless_ai/tests/test_ai_indexing.py
+++ b/src/paperless_ai/tests/test_ai_indexing.py
@@ -58,6 +58,26 @@ def test_build_document_node(real_document) -> None:
     assert nodes[0].metadata["document_id"] == str(real_document.id)
 
 
+@pytest.mark.django_db
+def test_build_document_node_excludes_metadata_from_embedding(real_document) -> None:
+    """Metadata keys must not be prepended to the embedding text.
+
+    build_llm_index_text already encodes all metadata in the body text, so
+    including it again via llama_index's default MetadataMode.EMBED would
+    double the token count and exceed embedding models with small context
+    windows (e.g. nomic-embed-text via Ollama defaults to num_ctx=2048).
+    """
+    from llama_index.core.schema import MetadataMode
+
+    nodes = indexing.build_document_node(real_document)
+    for node in nodes:
+        embed_text = node.get_content(metadata_mode=MetadataMode.EMBED)
+        for key in node.metadata:
+            assert key not in embed_text, (
+                f"Metadata key '{key}' should not appear in embedding text"
+            )
+
+
 @pytest.mark.django_db
 def test_build_document_node_uses_rag_chunk_settings(real_document) -> None:
     with patch("llama_index.core.node_parser.SimpleNodeParser") as mock_parser: