Feat: apply LLM to optimize citations. (#5935)

### What problem does this PR solve?

#5905

### Type of change

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

rag/prompts.py

@@ -108,22 +108,63 @@ def kb_prompt(kbinfos, max_tokens):
     docs = {d.id: d.meta_fields for d in docs}
 
     doc2chunks = defaultdict(lambda: {"chunks": [], "meta": []})
-    for ck in kbinfos["chunks"][:chunks_num]:
-        doc2chunks[ck["docnm_kwd"]]["chunks"].append((f"URL: {ck['url']}\n" if "url" in ck else "") + ck["content_with_weight"])
+    for i, ck in enumerate(kbinfos["chunks"][:chunks_num]):
+        doc2chunks[ck["docnm_kwd"]]["chunks"].append((f"URL: {ck['url']}\n" if "url" in ck else "") + f"ID: {i}\n" + ck["content_with_weight"])
         doc2chunks[ck["docnm_kwd"]]["meta"] = docs.get(ck["doc_id"], {})
 
     knowledges = []
     for nm, cks_meta in doc2chunks.items():
-        txt = f"Document: {nm} \n"
+        txt = f"\nDocument: {nm} \n"
         for k, v in cks_meta["meta"].items():
             txt += f"{k}: {v}\n"
         txt += "Relevant fragments as following:\n"
         for i, chunk in enumerate(cks_meta["chunks"], 1):
-            txt += f"{i}. {chunk}\n"
+            txt += f"{chunk}\n"
         knowledges.append(txt)
     return knowledges
 
 
+def citation_prompt():
+    return """
+
+# Citation requirements:
+- Inserts CITATIONS in format '##i$$ ##j$$' where i,j are the ID of the content you are citing and encapsulated with '##' and '$$'.
+- Inserts the CITATION symbols at the end of a sentence, AND NO MORE than 4 citations.
+- DO NOT insert CITATION in the answer if the content is not from retrieved chunks.
+
+--- Example START ---
+<SYSTEM>: Here is the knowledge base:
+
+Document: Elon Musk Breaks Silence on Crypto, Warns Against Dogecoin ...
+URL: https://blockworks.co/news/elon-musk-crypto-dogecoin
+ID: 0
+The Tesla co-founder advised against going all-in on dogecoin, but Elon Musk said it’s still his favorite crypto...
+
+Document: Elon Musk's Dogecoin tweet sparks social media frenzy
+ID: 1
+Musk said he is 'willing to serve' D.O.G.E. – shorthand for Dogecoin.
+
+Document: Causal effect of Elon Musk tweets on Dogecoin price
+ID: 2
+If you think of Dogecoin — the cryptocurrency based on a meme — you can’t help but also think of Elon Musk...
+
+Document: Elon Musk's Tweet Ignites Dogecoin's Future In Public Services
+ID: 3
+The market is heating up after Elon Musk's announcement about Dogecoin. Is this a new era for crypto?...
+
+      The above is the knowledge base.
+
+<USER>: What's the Elon's view on dogecoin?
+
+<ASSISTANT>: Musk has consistently expressed his fondness for Dogecoin, often citing its humor and the inclusion of dogs in its branding. He has referred to it as his favorite cryptocurrency ##0$$ ##1$$.
+Recently, Musk has hinted at potential future roles for Dogecoin. His tweets have sparked speculation about Dogecoin's potential integration into public services ##3$$.
+Overall, while Musk enjoys Dogecoin and often promotes it, he also warns against over-investing in it, reflecting both his personal amusement and caution regarding its speculative nature.
+
+--- Example END ---
+
+"""
+
+
 def keyword_extraction(chat_mdl, content, topn=3):
     prompt = f"""
 Role: You're a text analyzer.