Feat: support parent-child in search procedure. (#11629)

### What problem does this PR solve?

#7996

### Type of change

- [x] New Feature (non-breaking change which adds functionality)

rag/nlp/search.py

@@ -17,7 +17,7 @@ import json
 import logging
 import re
 import math
-from collections import OrderedDict
+from collections import OrderedDict, defaultdict
 from dataclasses import dataclass
 
 from rag.prompts.generator import relevant_chunks_with_toc
@@ -640,3 +640,50 @@ class Dealer:
             chunks.append(d)
 
         return sorted(chunks, key=lambda x:x["similarity"]*-1)[:topn]
+
+    def retrieval_by_children(self, chunks:list[dict], tenant_ids:list[str]):
+        if not chunks:
+            return []
+        idx_nms = [index_name(tid) for tid in tenant_ids]
+        mom_chunks = defaultdict([])
+        i = 0
+        while i < len(chunks):
+            ck = chunks[i]
+            if not ck.get("mom_id"):
+                i += 1
+                continue
+            mom_chunks[ck["mom_id"]].append(chunks.pop(i))
+
+        if not mom_chunks:
+            return chunks
+
+        if not chunks:
+            chunks = []
+
+        vector_size = 1024
+        for id, cks in mom_chunks.items():
+            chunk = self.dataStore.get(id, idx_nms, [ck["kb_id"] for ck in cks])
+            d = {
+                "chunk_id": id,
+                "content_ltks": " ".join([ck["content_ltks"] for ck in cks]),
+                "content_with_weight": chunk["content_with_weight"],
+                "doc_id": chunk["doc_id"],
+                "docnm_kwd": chunk.get("docnm_kwd", ""),
+                "kb_id": chunk["kb_id"],
+                "important_kwd": [kwd for ck in cks for kwd in ck.get("important_kwd", [])],
+                "image_id": chunk.get("img_id", ""),
+                "similarity": np.mean([ck["similarity"] for ck in cks]),
+                "vector_similarity": np.mean([ck["similarity"] for ck in cks]),
+                "term_similarity": np.mean([ck["similarity"] for ck in cks]),
+                "vector": [0.0] * vector_size,
+                "positions": chunk.get("position_int", []),
+                "doc_type_kwd": chunk.get("doc_type_kwd", "")
+            }
+            for k in cks[0].keys():
+                if k[-4:] == "_vec":
+                    d["vector"] = cks[0][k]
+                    vector_size = len(cks[0][k])
+                    break
+            chunks.append(d)
+
+        return sorted(chunks, key=lambda x:x["similarity"]*-1)