trim input to TGI, moved clustering and summarization to dataprep and store in DB

comps/cores/mega/gateway.py

@@ -1048,13 +1048,13 @@ def parser_input(data, TypeClass, key):
             if isinstance(response, StreamingResponse):
                 return response
         last_node = runtime_graph.all_leaves()[-1]
-        response = result_dict[last_node]["text"]
+        response_content = result_dict[last_node]["choices"][0]["message"]["content"]
         choices = []
         usage = UsageInfo()
         choices.append(
             ChatCompletionResponseChoice(
                 index=0,
-                message=ChatMessage(role="assistant", content=response),
+                message=ChatMessage(role="assistant", content=response_content),
                 finish_reason="stop",
             )
         )

comps/dataprep/neo4j/llama_index/README.md

@@ -1,5 +1,14 @@
 # Dataprep Microservice with Neo4J
 
+This Dataprep microservice performs:
+
+- Graph extraction (entities, relationships and descripttions) using LLM
+- Performs hierarchical_leiden clustering to identify communities in the knowledge graph
+- Generates a community symmary for each community
+- Stores all of the above in Neo4j Graph DB
+
+This miroservice follows the graphRAG approached defined by Microsoft paper "From Local to Global: A Graph RAG Approach to Query-Focused Summarization" with some differences such as: 1) only level zero cluster summaries are leveraged, 2) The input context to the final answer generation is trimmed to fit maximum context length.
+
 This dataprep microservice ingests the input files and uses LLM (TGI or OpenAI model when OPENAI_API_KEY is set) to extract entities, relationships and descriptions of those to build a graph-based text index.
 
 ## Setup Environment Variables
@@ -78,6 +87,11 @@ curl -X POST \
     http://${host_ip}:6004/v1/dataprep
 ```
 
+Please note that clustering of extracted entities and summarization happens in this data preparation step. The consecuence of this is:
+
+- Large processing time for large dataset. An LLM call is done to summarize each cluster which may result in large volume of LLM calls
+- Need to clean graph GB entity_info and Cluster if dataprep is run multiple times since the resulting cluster numbering will differ between consecutive calls and will corrupt the results.
+
 We support table extraction from pdf documents. You can specify process_table and table_strategy by the following commands. "table_strategy" refers to the strategies to understand tables for table retrieval. As the setting progresses from "fast" to "hq" to "llm," the focus shifts towards deeper table understanding at the expense of processing speed. The default strategy is "fast".
 
 Note: If you specify "table_strategy=llm" TGI service will be used.