Store docling-serve chunk bodies via raw_text

CHANGELOG.md

@@ -15,6 +15,7 @@
 ### Fixed
 
 - `docling-serve` converter requests `include_page_images`, so `generate_page_images` produces page rasters on docling-serve versions that gate them behind that flag.
+- `docling-serve` chunker stores chunk bodies via `raw_text`, so section headings aren't duplicated into chunk content on docling-serve versions where the chunk `text` field is heading-contextualized.
 
 ## [0.60.0] - 2026-06-22
 

docs/remote-processing.md

@@ -30,6 +30,8 @@ docling-serve is a REST API service that provides:
 
 ## Setup
 
+haiku.rag is tested against docling-serve 1.25.0.
+
 ### Docker Compose (Recommended)
 
 The slim Docker image with docker-compose is the recommended setup. See `examples/docker/docker-compose.yml` for a complete configuration that includes both services.

haiku_rag_slim/haiku/rag/chunkers/docling_serve.py

@@ -77,6 +77,7 @@ class DoclingServeChunker(DocumentChunker):
             "chunking_use_markdown_tables": str(
                 self.config.processing.chunking_use_markdown_tables
             ).lower(),
+            "chunking_include_raw_text": "true",
         }
         if opts.ocr_lang:
             data["convert_ocr_lang"] = opts.ocr_lang
@@ -154,7 +155,10 @@ class DoclingServeChunker(DocumentChunker):
         result: list[Chunk] = []
 
         for chunk in raw_chunks:
-            text = chunk.get("text", "")
+            # ``text`` is the embedding-targeted serialization (headings and
+            # captions prepended); ``raw_text`` is the bare body. Store the
+            # body and let the embedding/FTS layer contextualize from headings.
+            text = chunk.get("raw_text") or chunk.get("text", "")
 
             # doc_items from docling-serve is a list of ref strings like ["#/texts/1", "#/tables/0"]
             doc_items = chunk.get("doc_items", [])