Gurjot Singh
commited on
Commit
·
b0187f6
1
Parent(s):
c0af224
Add custom function with separate keyword extraction for user's query and a separate prompt
Browse files- lightrag/base.py +2 -0
- lightrag/lightrag.py +110 -0
- lightrag/operate.py +200 -0
- test.py +1 -1
lightrag/base.py
CHANGED
@@ -31,6 +31,8 @@ class QueryParam:
|
|
31 |
max_token_for_global_context: int = 4000
|
32 |
# Number of tokens for the entity descriptions
|
33 |
max_token_for_local_context: int = 4000
|
|
|
|
|
34 |
|
35 |
|
36 |
@dataclass
|
|
|
31 |
max_token_for_global_context: int = 4000
|
32 |
# Number of tokens for the entity descriptions
|
33 |
max_token_for_local_context: int = 4000
|
34 |
+
hl_keywords: list[str] = field(default_factory=list)
|
35 |
+
ll_keywords: list[str] = field(default_factory=list)
|
36 |
|
37 |
|
38 |
@dataclass
|
lightrag/lightrag.py
CHANGED
@@ -17,6 +17,8 @@ from .operate import (
|
|
17 |
kg_query,
|
18 |
naive_query,
|
19 |
mix_kg_vector_query,
|
|
|
|
|
20 |
)
|
21 |
|
22 |
from .utils import (
|
@@ -753,6 +755,114 @@ class LightRAG:
|
|
753 |
await self._query_done()
|
754 |
return response
|
755 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
756 |
async def _query_done(self):
|
757 |
tasks = []
|
758 |
for storage_inst in [self.llm_response_cache]:
|
|
|
17 |
kg_query,
|
18 |
naive_query,
|
19 |
mix_kg_vector_query,
|
20 |
+
extract_keywords_only,
|
21 |
+
kg_query_with_keywords,
|
22 |
)
|
23 |
|
24 |
from .utils import (
|
|
|
755 |
await self._query_done()
|
756 |
return response
|
757 |
|
758 |
+
def query_with_separate_keyword_extraction(
|
759 |
+
self,
|
760 |
+
query: str,
|
761 |
+
prompt: str,
|
762 |
+
param: QueryParam = QueryParam()
|
763 |
+
):
|
764 |
+
"""
|
765 |
+
1. Extract keywords from the 'query' using new function in operate.py.
|
766 |
+
2. Then run the standard aquery() flow with the final prompt (formatted_question).
|
767 |
+
"""
|
768 |
+
|
769 |
+
loop = always_get_an_event_loop()
|
770 |
+
return loop.run_until_complete(self.aquery_with_separate_keyword_extraction(query, prompt, param))
|
771 |
+
|
772 |
+
async def aquery_with_separate_keyword_extraction(
|
773 |
+
self,
|
774 |
+
query: str,
|
775 |
+
prompt: str,
|
776 |
+
param: QueryParam = QueryParam()
|
777 |
+
):
|
778 |
+
"""
|
779 |
+
1. Calls extract_keywords_only to get HL/LL keywords from 'query'.
|
780 |
+
2. Then calls kg_query(...) or naive_query(...), etc. as the main query, while also injecting the newly extracted keywords if needed.
|
781 |
+
"""
|
782 |
+
|
783 |
+
# ---------------------
|
784 |
+
# STEP 1: Keyword Extraction
|
785 |
+
# ---------------------
|
786 |
+
# We'll assume 'extract_keywords_only(...)' returns (hl_keywords, ll_keywords).
|
787 |
+
hl_keywords, ll_keywords = await extract_keywords_only(
|
788 |
+
text=query,
|
789 |
+
param=param,
|
790 |
+
global_config=asdict(self),
|
791 |
+
hashing_kv=self.llm_response_cache or self.key_string_value_json_storage_cls(
|
792 |
+
namespace="llm_response_cache",
|
793 |
+
global_config=asdict(self),
|
794 |
+
embedding_func=None,
|
795 |
+
)
|
796 |
+
)
|
797 |
+
|
798 |
+
param.hl_keywords=hl_keywords,
|
799 |
+
param.ll_keywords=ll_keywords,
|
800 |
+
|
801 |
+
# ---------------------
|
802 |
+
# STEP 2: Final Query Logic
|
803 |
+
# ---------------------
|
804 |
+
|
805 |
+
# Create a new string with the prompt and the keywords
|
806 |
+
ll_keywords_str = ", ".join(ll_keywords)
|
807 |
+
hl_keywords_str = ", ".join(hl_keywords)
|
808 |
+
formatted_question = f"{prompt}\n\n### Keywords:\nHigh-level: {hl_keywords_str}\nLow-level: {ll_keywords_str}\n\n### Query:\n{query}"
|
809 |
+
|
810 |
+
if param.mode in ["local", "global", "hybrid"]:
|
811 |
+
response = await kg_query_with_keywords(
|
812 |
+
formatted_question,
|
813 |
+
self.chunk_entity_relation_graph,
|
814 |
+
self.entities_vdb,
|
815 |
+
self.relationships_vdb,
|
816 |
+
self.text_chunks,
|
817 |
+
param,
|
818 |
+
asdict(self),
|
819 |
+
hashing_kv=self.llm_response_cache
|
820 |
+
if self.llm_response_cache and hasattr(self.llm_response_cache, "global_config")
|
821 |
+
else self.key_string_value_json_storage_cls(
|
822 |
+
namespace="llm_response_cache",
|
823 |
+
global_config=asdict(self),
|
824 |
+
embedding_func=None,
|
825 |
+
),
|
826 |
+
)
|
827 |
+
elif param.mode == "naive":
|
828 |
+
response = await naive_query(
|
829 |
+
formatted_question,
|
830 |
+
self.chunks_vdb,
|
831 |
+
self.text_chunks,
|
832 |
+
param,
|
833 |
+
asdict(self),
|
834 |
+
hashing_kv=self.llm_response_cache
|
835 |
+
if self.llm_response_cache and hasattr(self.llm_response_cache, "global_config")
|
836 |
+
else self.key_string_value_json_storage_cls(
|
837 |
+
namespace="llm_response_cache",
|
838 |
+
global_config=asdict(self),
|
839 |
+
embedding_func=None,
|
840 |
+
),
|
841 |
+
)
|
842 |
+
elif param.mode == "mix":
|
843 |
+
response = await mix_kg_vector_query(
|
844 |
+
formatted_question,
|
845 |
+
self.chunk_entity_relation_graph,
|
846 |
+
self.entities_vdb,
|
847 |
+
self.relationships_vdb,
|
848 |
+
self.chunks_vdb,
|
849 |
+
self.text_chunks,
|
850 |
+
param,
|
851 |
+
asdict(self),
|
852 |
+
hashing_kv=self.llm_response_cache
|
853 |
+
if self.llm_response_cache and hasattr(self.llm_response_cache, "global_config")
|
854 |
+
else self.key_string_value_json_storage_cls(
|
855 |
+
namespace="llm_response_cache",
|
856 |
+
global_config=asdict(self),
|
857 |
+
embedding_func=None,
|
858 |
+
),
|
859 |
+
)
|
860 |
+
else:
|
861 |
+
raise ValueError(f"Unknown mode {param.mode}")
|
862 |
+
|
863 |
+
await self._query_done()
|
864 |
+
return response
|
865 |
+
|
866 |
async def _query_done(self):
|
867 |
tasks = []
|
868 |
for storage_inst in [self.llm_response_cache]:
|
lightrag/operate.py
CHANGED
@@ -680,6 +680,206 @@ async def kg_query(
|
|
680 |
)
|
681 |
return response
|
682 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
683 |
|
684 |
async def _build_query_context(
|
685 |
query: list,
|
|
|
680 |
)
|
681 |
return response
|
682 |
|
683 |
+
async def kg_query_with_keywords(
|
684 |
+
query: str,
|
685 |
+
knowledge_graph_inst: BaseGraphStorage,
|
686 |
+
entities_vdb: BaseVectorStorage,
|
687 |
+
relationships_vdb: BaseVectorStorage,
|
688 |
+
text_chunks_db: BaseKVStorage[TextChunkSchema],
|
689 |
+
query_param: QueryParam,
|
690 |
+
global_config: dict,
|
691 |
+
hashing_kv: BaseKVStorage = None,
|
692 |
+
) -> str:
|
693 |
+
"""
|
694 |
+
Refactored kg_query that does NOT extract keywords by itself.
|
695 |
+
It expects hl_keywords and ll_keywords to be set in query_param, or defaults to empty.
|
696 |
+
Then it uses those to build context and produce a final LLM response.
|
697 |
+
"""
|
698 |
+
|
699 |
+
# ---------------------------
|
700 |
+
# 0) Handle potential cache
|
701 |
+
# ---------------------------
|
702 |
+
use_model_func = global_config["llm_model_func"]
|
703 |
+
args_hash = compute_args_hash(query_param.mode, query)
|
704 |
+
cached_response, quantized, min_val, max_val = await handle_cache(
|
705 |
+
hashing_kv, args_hash, query, query_param.mode
|
706 |
+
)
|
707 |
+
if cached_response is not None:
|
708 |
+
return cached_response
|
709 |
+
|
710 |
+
# ---------------------------
|
711 |
+
# 1) RETRIEVE KEYWORDS FROM query_param
|
712 |
+
# ---------------------------
|
713 |
+
|
714 |
+
# If these fields don't exist, default to empty lists/strings.
|
715 |
+
hl_keywords = getattr(query_param, "hl_keywords", []) or []
|
716 |
+
ll_keywords = getattr(query_param, "ll_keywords", []) or []
|
717 |
+
|
718 |
+
# If neither has any keywords, you could handle that logic here.
|
719 |
+
if not hl_keywords and not ll_keywords:
|
720 |
+
logger.warning("No keywords found in query_param. Could default to global mode or fail.")
|
721 |
+
return PROMPTS["fail_response"]
|
722 |
+
if not ll_keywords and query_param.mode in ["local", "hybrid"]:
|
723 |
+
logger.warning("low_level_keywords is empty, switching to global mode.")
|
724 |
+
query_param.mode = "global"
|
725 |
+
if not hl_keywords and query_param.mode in ["global", "hybrid"]:
|
726 |
+
logger.warning("high_level_keywords is empty, switching to local mode.")
|
727 |
+
query_param.mode = "local"
|
728 |
+
|
729 |
+
# Flatten low-level and high-level keywords if needed
|
730 |
+
ll_keywords_flat = [item for sublist in ll_keywords for item in sublist] if any(isinstance(i, list) for i in ll_keywords) else ll_keywords
|
731 |
+
hl_keywords_flat = [item for sublist in hl_keywords for item in sublist] if any(isinstance(i, list) for i in hl_keywords) else hl_keywords
|
732 |
+
|
733 |
+
# Join the flattened lists
|
734 |
+
ll_keywords_str = ", ".join(ll_keywords_flat) if ll_keywords_flat else ""
|
735 |
+
hl_keywords_str = ", ".join(hl_keywords_flat) if hl_keywords_flat else ""
|
736 |
+
|
737 |
+
keywords = [ll_keywords_str, hl_keywords_str]
|
738 |
+
|
739 |
+
logger.info("Using %s mode for query processing", query_param.mode)
|
740 |
+
|
741 |
+
# ---------------------------
|
742 |
+
# 2) BUILD CONTEXT
|
743 |
+
# ---------------------------
|
744 |
+
context = await _build_query_context(
|
745 |
+
keywords,
|
746 |
+
knowledge_graph_inst,
|
747 |
+
entities_vdb,
|
748 |
+
relationships_vdb,
|
749 |
+
text_chunks_db,
|
750 |
+
query_param,
|
751 |
+
)
|
752 |
+
if not context:
|
753 |
+
return PROMPTS["fail_response"]
|
754 |
+
|
755 |
+
# If only context is needed, return it
|
756 |
+
if query_param.only_need_context:
|
757 |
+
return context
|
758 |
+
|
759 |
+
# ---------------------------
|
760 |
+
# 3) BUILD THE SYSTEM PROMPT + CALL LLM
|
761 |
+
# ---------------------------
|
762 |
+
sys_prompt_temp = PROMPTS["rag_response"]
|
763 |
+
sys_prompt = sys_prompt_temp.format(
|
764 |
+
context_data=context, response_type=query_param.response_type
|
765 |
+
)
|
766 |
+
|
767 |
+
if query_param.only_need_prompt:
|
768 |
+
return sys_prompt
|
769 |
+
|
770 |
+
# Now call the LLM with the final system prompt
|
771 |
+
response = await use_model_func(
|
772 |
+
query,
|
773 |
+
system_prompt=sys_prompt,
|
774 |
+
stream=query_param.stream,
|
775 |
+
)
|
776 |
+
|
777 |
+
# Clean up the response
|
778 |
+
if isinstance(response, str) and len(response) > len(sys_prompt):
|
779 |
+
response = (
|
780 |
+
response.replace(sys_prompt, "")
|
781 |
+
.replace("user", "")
|
782 |
+
.replace("model", "")
|
783 |
+
.replace(query, "")
|
784 |
+
.replace("<system>", "")
|
785 |
+
.replace("</system>", "")
|
786 |
+
.strip()
|
787 |
+
)
|
788 |
+
|
789 |
+
# ---------------------------
|
790 |
+
# 4) SAVE TO CACHE
|
791 |
+
# ---------------------------
|
792 |
+
await save_to_cache(
|
793 |
+
hashing_kv,
|
794 |
+
CacheData(
|
795 |
+
args_hash=args_hash,
|
796 |
+
content=response,
|
797 |
+
prompt=query,
|
798 |
+
quantized=quantized,
|
799 |
+
min_val=min_val,
|
800 |
+
max_val=max_val,
|
801 |
+
mode=query_param.mode,
|
802 |
+
),
|
803 |
+
)
|
804 |
+
return response
|
805 |
+
|
806 |
+
async def extract_keywords_only(
|
807 |
+
text: str,
|
808 |
+
param: QueryParam,
|
809 |
+
global_config: dict,
|
810 |
+
hashing_kv: BaseKVStorage = None,
|
811 |
+
) -> tuple[list[str], list[str]]:
|
812 |
+
"""
|
813 |
+
Extract high-level and low-level keywords from the given 'text' using the LLM.
|
814 |
+
This method does NOT build the final RAG context or provide a final answer.
|
815 |
+
It ONLY extracts keywords (hl_keywords, ll_keywords).
|
816 |
+
"""
|
817 |
+
|
818 |
+
# 1. Handle cache if needed
|
819 |
+
args_hash = compute_args_hash(param.mode, text)
|
820 |
+
cached_response, quantized, min_val, max_val = await handle_cache(
|
821 |
+
hashing_kv, args_hash, text, param.mode
|
822 |
+
)
|
823 |
+
if cached_response is not None:
|
824 |
+
# parse the cached_response if it’s JSON containing keywords
|
825 |
+
# or simply return (hl_keywords, ll_keywords) from cached
|
826 |
+
# Assuming cached_response is in the same JSON structure:
|
827 |
+
match = re.search(r"\{.*\}", cached_response, re.DOTALL)
|
828 |
+
if match:
|
829 |
+
keywords_data = json.loads(match.group(0))
|
830 |
+
hl_keywords = keywords_data.get("high_level_keywords", [])
|
831 |
+
ll_keywords = keywords_data.get("low_level_keywords", [])
|
832 |
+
return hl_keywords, ll_keywords
|
833 |
+
return [], []
|
834 |
+
|
835 |
+
# 2. Build the examples
|
836 |
+
example_number = global_config["addon_params"].get("example_number", None)
|
837 |
+
if example_number and example_number < len(PROMPTS["keywords_extraction_examples"]):
|
838 |
+
examples = "\n".join(
|
839 |
+
PROMPTS["keywords_extraction_examples"][: int(example_number)]
|
840 |
+
)
|
841 |
+
else:
|
842 |
+
examples = "\n".join(PROMPTS["keywords_extraction_examples"])
|
843 |
+
language = global_config["addon_params"].get(
|
844 |
+
"language", PROMPTS["DEFAULT_LANGUAGE"]
|
845 |
+
)
|
846 |
+
|
847 |
+
# 3. Build the keyword-extraction prompt
|
848 |
+
kw_prompt_temp = PROMPTS["keywords_extraction"]
|
849 |
+
kw_prompt = kw_prompt_temp.format(query=text, examples=examples, language=language)
|
850 |
+
|
851 |
+
# 4. Call the LLM for keyword extraction
|
852 |
+
use_model_func = global_config["llm_model_func"]
|
853 |
+
result = await use_model_func(kw_prompt, keyword_extraction=True)
|
854 |
+
|
855 |
+
# 5. Parse out JSON from the LLM response
|
856 |
+
match = re.search(r"\{.*\}", result, re.DOTALL)
|
857 |
+
if not match:
|
858 |
+
logger.error("No JSON-like structure found in the result.")
|
859 |
+
return [], []
|
860 |
+
try:
|
861 |
+
keywords_data = json.loads(match.group(0))
|
862 |
+
except json.JSONDecodeError as e:
|
863 |
+
logger.error(f"JSON parsing error: {e}")
|
864 |
+
return [], []
|
865 |
+
|
866 |
+
hl_keywords = keywords_data.get("high_level_keywords", [])
|
867 |
+
ll_keywords = keywords_data.get("low_level_keywords", [])
|
868 |
+
|
869 |
+
# 6. Cache the result if needed
|
870 |
+
await save_to_cache(
|
871 |
+
hashing_kv,
|
872 |
+
CacheData(
|
873 |
+
args_hash=args_hash,
|
874 |
+
content=result,
|
875 |
+
prompt=text,
|
876 |
+
quantized=quantized,
|
877 |
+
min_val=min_val,
|
878 |
+
max_val=max_val,
|
879 |
+
mode=param.mode,
|
880 |
+
),
|
881 |
+
)
|
882 |
+
return hl_keywords, ll_keywords
|
883 |
|
884 |
async def _build_query_context(
|
885 |
query: list,
|
test.py
CHANGED
@@ -39,4 +39,4 @@ print(
|
|
39 |
# Perform hybrid search
|
40 |
print(
|
41 |
rag.query("What are the top themes in this story?", param=QueryParam(mode="hybrid"))
|
42 |
-
)
|
|
|
39 |
# Perform hybrid search
|
40 |
print(
|
41 |
rag.query("What are the top themes in this story?", param=QueryParam(mode="hybrid"))
|
42 |
+
)
|