Update FM streaming RAG demo to use new NIMs (NVIDIA#234)

* Update FM streaming RAG demo for DC AI Summit

- Uses new embedding and reranking NIMs
- Uses Parakeet ASR NIM
- Includes GNU Radio container
- Updates deployment scripts for deploying across multiple machines
- Various bug fixes and QoL updates

* Add diagramss for README

* Add README

* By default, use build.nvidia.com endpoints for non-ASR NIMs

* Restart Riva gRPC service on exceptions

* Add GNU Radio folder

* Update README to include frontend URL

* Fix LLM model options

* Fix bug when not using local Embedding / Reranking models

Author: dylan-eustice

community/fm-asr-streaming-rag/.gitignore

@@ -0,0 +1,2 @@
+deploy/.keys
+milvus/volumes

community/fm-asr-streaming-rag/README.md

@@ -13,12 +13,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
+import datetime
+import requests
 from common import get_logger
 from database import TimestampDatabase
 from langchain.text_splitter import RecursiveCharacterTextSplitter
 
 logger = get_logger(__name__)
 
+FRONTEND_URI = os.environ.get('FRONTEND_URI', None)
+
 #todo: Multi-thread to handle multiple concurrent streams
 #todo: Add time-triggered embedding (i.e. embed after N seconds if no updates)
 class TextAccumulator:
@@ -45,4 +50,14 @@ def update(self, source_id, text):
         self.timestamp_db.insert_docs(new_docs, source_id)
         self.db_interface.add_docs(new_docs, source_id)
 
+        for doc in new_docs:
+            endpoint = f"http://{FRONTEND_URI}/app/update_finalized_transcript"
+            time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            try:
+                client_response = requests.post(endpoint, json={'transcript': f"[{time}] {doc}"})
+                logger.debug(f'Posted update_finalized_transcript: {client_response._content}')
+                logger.debug("--------------------------")
+            except requests.exceptions.ConnectionError:
+                logger.error(f"Failed to connect to the '{endpoint}' endpoint")
+
         return {"status": f"Added {len(new_docs)} entries"}

community/fm-asr-streaming-rag/chain-server/accumulator.py

@@ -13,15 +13,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Union
 from copy import copy
 from datetime import datetime, timedelta
 from langchain_core.output_parsers import StrOutputParser
 from langchain_core.prompts import ChatPromptTemplate
 from langchain.docstore.document import Document
 
 from accumulator import TextAccumulator
-from retriever import NemoRetrieverInterface, NvidiaApiInterface
+from retriever import NVRetriever
 from common import get_logger, LLMConfig, TimeResponse, UserIntent
 from utils import get_llm, classify
 from prompts import RAG_PROMPT, INTENT_PROMPT, RECENCY_PROMPT, SUMMARIZATION_PROMPT
@@ -36,7 +35,7 @@ def __init__(
         self,
         config: LLMConfig,
         text_accumulator: TextAccumulator,
-        retv_interface: Union[NemoRetrieverInterface, NvidiaApiInterface]
+        retv_interface: NVRetriever
     ):
         self.config = config
         self.text_accumulator = text_accumulator

community/fm-asr-streaming-rag/chain-server/chains.py

@@ -24,8 +24,13 @@
 from typing import Literal
 from langchain_community.utils.math import cosine_similarity
 
-USE_NEMO_RETRIEVER = os.environ.get('USE_NEMO_RETRIEVER', 'False').lower() in ('true', '1')
-NVIDIA_API_KEY = os.environ.get('NVIDIA_API_KEY', 'null')
+NVIDIA_API_KEY  = os.environ.get('NVIDIA_API_KEY', 'null')
+LLM_URI         = os.environ.get('LLM_URI', None)
+RERANKING_MODEL = os.environ.get('RERANK_MODEL', None)
+RERANKING_URI   = os.environ.get('RERANK_URI', None)
+EMBEDDING_MODEL = os.environ.get('EMBED_MODEL', None)
+EMBEDDING_URI   = os.environ.get('EMBED_URI', None)
+MAX_DOCS        = int(os.environ.get('MAX_DOCS_RETR', 8))
 
 def get_logger(name):
     LOG_LEVEL = logging.getLevelName(os.environ.get('CHAIN_LOG_LEVEL', 'WARN').upper())
@@ -51,7 +56,7 @@ class LLMConfig(BaseModel):
     )
     # Model choice
     name: str = Field("Name of LLM instance to use")
-    engine: str = Field("Name of engine ['nvai-api-endpoint', 'triton-trt-llm']")
+    engine: str = Field("Name of engine ['nvai-api-endpoint', 'local-nim']")
     # Chain parameters
     use_knowledge_base: bool = Field(
         description="Whether to use a knowledge base", default=True
@@ -95,11 +100,7 @@ def nvapi_embedding(text):
     return embeddings
 
 VALID_TIME_UNITS = ["seconds", "minutes", "hours", "days"]
-TIME_VECTORS = None # Lazy loading in 'sanitize_time_unit'
-if USE_NEMO_RETRIEVER:
-    embedding_service = nemo_embedding
-else:
-    embedding_service = nvapi_embedding
+TIME_VECTORS = nvapi_embedding(VALID_TIME_UNITS)
 
 def sanitize_time_unit(time_unit):
     """
@@ -111,10 +112,7 @@ def sanitize_time_unit(time_unit):
     if time_unit in VALID_TIME_UNITS:
         return time_unit
 
-    if TIME_VECTORS is None:
-        TIME_VECTORS = embedding_service(VALID_TIME_UNITS)
-
-    unit_embedding = embedding_service([time_unit])
+    unit_embedding = nvapi_embedding([time_unit])
     similarity = cosine_similarity(unit_embedding, TIME_VECTORS)
     return VALID_TIME_UNITS[np.argmax(similarity)]
 

community/fm-asr-streaming-rag/chain-server/common.py

@@ -30,7 +30,7 @@
 import numpy as np
 
 from common import get_logger
-from datetime import datetime
+from datetime import datetime, timedelta
 from langchain.docstore.document import Document
 
 logger = get_logger(__name__)
@@ -87,8 +87,8 @@ def recent(self, tstamp):
     def past(self, tstamp, window=90):
         """ Return entries within 'window' seconds of tstamp
         """
-        tstart = tstamp - datetime.timedelta(seconds=window)
-        tend = tstamp + datetime.timedelta(seconds=window)
+        tstart = tstamp - timedelta(seconds=window)
+        tend = tstamp + timedelta(seconds=window)
         self.cursor.execute('SELECT * FROM messages WHERE timestamp BETWEEN ? AND ?', (tstart, tend))
         docs = self.cursor.fetchall()
         return [self.reformat(doc) for doc in docs]

community/fm-asr-streaming-rag/chain-server/database.py

@@ -49,7 +49,7 @@ def format_json(text: str):
 recent timeframe. Examples: "Can you summarize the last hour of content?", "What have the main \
 topics been over the last 5 minutes?", "Tell me the main stories of the past 2 hours.".
 - 'TimeWindow': If the user is asking about the focus of the conversation from a specified time in \
-the past. Examples: "What were they talking about 15 minutes ago?", "What was the focus an hour ago?".
+the past. Examples: "What were they talking about 15 minutes ago?", "What was the focus an hour ago?", "What are they talking about right now?".
 - If the user's intent is not clear, or if the intent cannot be confidently determined, classify \
 this as 'Unknown'.
 
@@ -101,11 +101,17 @@ def format_json(text: str):
 f"'{recency_examples[1]}' --> '{format_json(recency_examples_obj[1].model_dump_json())}'\n" +
 f"'{recency_examples[2]}' --> '{format_json(recency_examples_obj[2].model_dump_json())}'\n" + """
 
-Convert the user input below into this JSON format.
+Convert the user input below into this JSON format. Make sure you use valid JSON and \
+don't worry about escaping quotes, just give valid JSON blobs.
 """)
 
 SUMMARIZATION_PROMPT = """\
 You are a sophisticated summarization tool designed to condense large blocks \
 of text into a concise summary. Given the user text, reduce the character \
 count by distilling into only the most important information.
+
+Do not say you are summarizing, i.e. do not say "Here's a summary...", just \
+condense the text to the best of your abilities.
+
+Summary:
 """

community/fm-asr-streaming-rag/chain-server/prompts.py

@@ -1,21 +1,6 @@
 fastapi==0.104.1
 uvicorn[standard]==0.24.0
-python-multipart==0.0.6
-unstructured[all-docs]==0.11.2
-sentence-transformers==2.2.2
-llama-index==0.9.22
 pymilvus==2.3.5
-dataclass-wizard==0.22.2
-opencv-python==4.8.0.74
-minio==7.2.0
-asyncpg==0.29.0
-psycopg2-binary==2.9.9
-pgvector==0.2.4
-langchain==0.1.14
-langchain-core==0.1.40
-langchain-nvidia-ai-endpoints==0.0.12
-langchain-nvidia-trt==0.0.1rc0
-nemollm==0.3.4
-opentelemetry-sdk==1.21.0
-opentelemetry-api==1.21.0
-opentelemetry-exporter-otlp-proto-grpc==1.21.0
+langchain==0.2.6
+langchain_core==0.2.25
+langchain_nvidia_ai_endpoints==0.2.0