wehub-resource-sync
a0c8464e58
Build Package / build (ubuntu-latest) (push) Failing after 1s
CodeQL / Analyze (python) (push) Failing after 1s
Core Typecheck / core-typecheck (push) Failing after 1s
Linting / lint (push) Failing after 1s
llama-dev tests / test-llama-dev (push) Failing after 1s
Publish Sub-Package to PyPI if Needed / publish_subpackage_if_needed (push) Has been skipped
Sync Docs to Developer Hub / sync-docs (push) Failing after 0s
Build Package / build (windows-latest) (push) Has been cancelled
798 lines
30 KiB
Plaintext
798 lines
30 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"<a href=\"https://colab.research.google.com/github/run-llama/llama_index/blob/main/docs/examples/evaluation/Cleanlab.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"# Trustworthy RAG with LlamaIndex and Cleanlab\n",
|
||
"\n",
|
||
"LLMs occasionally hallucinate incorrect answers, especially for questions not well-supported within their training data. While organizations are adopting Retrieval Augmented Generation (RAG) to power LLMs with proprietary data, incorrect RAG responses remain a problem.\n",
|
||
"\n",
|
||
"This tutorial shows how to build **trustworthy** RAG applications: use [Cleanlab](https://help.cleanlab.ai/tlm/) to score the trustworthiness of every LLM response, and diagnose *why* responses are untrustworthy via evaluations of specific RAG components.\n",
|
||
"\n",
|
||
"Powered by [state-of-the-art uncertainty estimation](https://cleanlab.ai/blog/trustworthy-language-model/), Cleanlab trustworthiness scores help you automatically catch incorrect responses from any LLM application. Trust scoring happens in real-time and does not require any data labeling or model training work. Cleanlab provides additional real-time Evals for specific RAG components like the retrieved context, which help you root cause *why* RAG responses were incorrect. Cleanlab makes it easy to prevent inaccurate responses from your RAG app, and avoid losing your users' trust."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Setup\n",
|
||
"\n",
|
||
"This tutorial requires a:\n",
|
||
"- Cleanlab API Key: Sign up at [tlm.cleanlab.ai/](https://tlm.cleanlab.ai/) to get a free key\n",
|
||
"- OpenAI API Key: To make completion requests to an LLM\n",
|
||
"\n",
|
||
"Start by installing the required dependencies. "
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"%pip install llama-index cleanlab-tlm"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import os, re\n",
|
||
"from typing import List, ClassVar\n",
|
||
"import pandas as pd\n",
|
||
"\n",
|
||
"from llama_index.llms.openai import OpenAI\n",
|
||
"from llama_index.embeddings.openai import OpenAIEmbedding\n",
|
||
"\n",
|
||
"from cleanlab_tlm import TrustworthyRAG, Eval, get_default_evals"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"Initialize the OpenAI client using its API key."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"os.environ[\"OPENAI_API_KEY\"] = \"<your-openai-api-key>\"\n",
|
||
"\n",
|
||
"llm = OpenAI(model=\"gpt-4o-mini\")\n",
|
||
"embed_model = OpenAIEmbedding(embed_batch_size=10)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"Now, we initialize Cleanlab's client with default configurations. You can achieve better detection accuracy and latency by adjusting [optional configurations](https://help.cleanlab.ai/tlm/tutorials/tlm_advanced/)."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"os.environ[\"CLEANLAB_TLM_API_KEY\"] = \"<your-cleanlab-api-key\"\n",
|
||
"\n",
|
||
"trustworthy_rag = (\n",
|
||
" TrustworthyRAG()\n",
|
||
") # Optional configurations can improve accuracy/latency"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Read data\n",
|
||
"\n",
|
||
"This tutorial uses Nvidia’s Q1 FY2024 earnings report as an example data source for populating the RAG application's knowledge base."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"--2025-05-07 16:13:28-- https://cleanlab-public.s3.amazonaws.com/Datasets/NVIDIA_Financial_Results_Q1_FY2024.md\n",
|
||
"Resolving cleanlab-public.s3.amazonaws.com (cleanlab-public.s3.amazonaws.com)... 54.231.236.193, 16.182.70.65, 52.217.14.204, ...\n",
|
||
"Connecting to cleanlab-public.s3.amazonaws.com (cleanlab-public.s3.amazonaws.com)|54.231.236.193|:443... connected.\n",
|
||
"HTTP request sent, awaiting response... 200 OK\n",
|
||
"Length: 7379 (7.2K) [binary/octet-stream]\n",
|
||
"Saving to: ‘NVIDIA_Financial_Results_Q1_FY2024.md’\n",
|
||
"\n",
|
||
"NVIDIA_Financial_Re 100%[===================>] 7.21K --.-KB/s in 0s \n",
|
||
"\n",
|
||
"2025-05-07 16:13:28 (97.7 MB/s) - ‘NVIDIA_Financial_Results_Q1_FY2024.md’ saved [7379/7379]\n",
|
||
"\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"!wget -nc 'https://cleanlab-public.s3.amazonaws.com/Datasets/NVIDIA_Financial_Results_Q1_FY2024.md'\n",
|
||
"!mkdir -p ./data\n",
|
||
"!mv NVIDIA_Financial_Results_Q1_FY2024.md data/"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago \n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"with open(\n",
|
||
" \"data/NVIDIA_Financial_Results_Q1_FY2024.md\", \"r\", encoding=\"utf-8\"\n",
|
||
") as file:\n",
|
||
" data = file.read()\n",
|
||
"\n",
|
||
"print(data[:200])"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Build a RAG pipeline\n",
|
||
"\n",
|
||
"Now let's build a simple RAG pipeline with LlamaIndex. We have already initialized the OpenAI API for both as LLM and Embedding model."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"from llama_index.core import Settings, VectorStoreIndex, SimpleDirectoryReader\n",
|
||
"\n",
|
||
"Settings.llm = llm\n",
|
||
"Settings.embed_model = embed_model"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"### Load Data and Create Index + Query Engine\n",
|
||
"\n",
|
||
"Let's create an index from the document we just pulled above. We stick with the default index from LlamaIndex for this tutorial."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"documents = SimpleDirectoryReader(\"data\").load_data()\n",
|
||
"# Optional step since we're loading just one data file\n",
|
||
"for doc in documents:\n",
|
||
" doc.excluded_llm_metadata_keys.append(\n",
|
||
" \"file_path\"\n",
|
||
" ) # file_path wouldn't be a useful metadata to add to LLM's context since our datasource contains just 1 file\n",
|
||
"index = VectorStoreIndex.from_documents(documents)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"The generated index is used to power a query engine over the data."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"query_engine = index.as_query_engine()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"Note that Cleanlab is agnostic to the index and the query engine used for RAG, and is compatible with any choices you make for these components of your system.\n",
|
||
"\n",
|
||
"In addition, you can just use Cleanlab in an existing custom-built RAG pipeline (using any other LLM generator, streaming or not). <br>\n",
|
||
"Cleanlab just needs the prompt sent to your LLM (including system instructions, retrieved context, user query, etc.) and the generated response.\n",
|
||
"\n",
|
||
"We define an event handler that stores the prompt that LlamaIndex sends to the LLM. Refer to the [instrumentation documentation](https://docs.llamaindex.ai/en/stable/examples/instrumentation/basic_usage/) for more details."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"from llama_index.core.instrumentation import get_dispatcher\n",
|
||
"from llama_index.core.instrumentation.events import BaseEvent\n",
|
||
"from llama_index.core.instrumentation.event_handlers import BaseEventHandler\n",
|
||
"from llama_index.core.instrumentation.events.llm import LLMPredictStartEvent\n",
|
||
"\n",
|
||
"\n",
|
||
"class PromptEventHandler(BaseEventHandler):\n",
|
||
" events: ClassVar[List[BaseEvent]] = []\n",
|
||
" PROMPT_TEMPLATE: str = \"\"\n",
|
||
"\n",
|
||
" @classmethod\n",
|
||
" def class_name(cls) -> str:\n",
|
||
" return \"PromptEventHandler\"\n",
|
||
"\n",
|
||
" def handle(self, event) -> None:\n",
|
||
" if isinstance(event, LLMPredictStartEvent):\n",
|
||
" self.PROMPT_TEMPLATE = event.template.default_template.template\n",
|
||
" self.events.append(event)\n",
|
||
"\n",
|
||
"\n",
|
||
"# Root dispatcher\n",
|
||
"root_dispatcher = get_dispatcher()\n",
|
||
"\n",
|
||
"# Register event handler\n",
|
||
"event_handler = PromptEventHandler()\n",
|
||
"root_dispatcher.add_event_handler(event_handler)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"For each query, we can fetch the prompt from `event_handler.PROMPT_TEMPLATE`. Let's see it in action."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Use our RAG application\n",
|
||
"\n",
|
||
"Now that the vector database is loaded with text chunks and their corresponding embeddings, we can start querying it to answer questions."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"NVIDIA's total revenue in the first quarter of fiscal 2024 was $7.19 billion.\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"query = \"What was NVIDIA's total revenue in the first quarter of fiscal 2024?\"\n",
|
||
"\n",
|
||
"response = query_engine.query(query)\n",
|
||
"print(response)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"This response is indeed correct for our simple query. Let's see the document chunks that LlamaIndex retrieved for this query, from which we can easy verify this response was right."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"def get_retrieved_context(response, print_chunks=False):\n",
|
||
" if isinstance(response, list):\n",
|
||
" texts = [node.text for node in response]\n",
|
||
" else:\n",
|
||
" texts = [src.node.text for src in response.source_nodes]\n",
|
||
"\n",
|
||
" if print_chunks:\n",
|
||
" for idx, text in enumerate(texts):\n",
|
||
" print(f\"--- Chunk {idx + 1} ---\\n{text[:200]}...\")\n",
|
||
" return \"\\n\".join(texts)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"--- Chunk 1 ---\n",
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago ...\n",
|
||
"--- Chunk 2 ---\n",
|
||
"- **Gross Margins**: GAAP and non-GAAP gross margins are expected to be 68.6% and 70.0%, respectively, plus or minus 50 basis points.\n",
|
||
"- **Operating Expenses**: GAAP and non-GAAP operating expenses are...\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"context_str = get_retrieved_context(response, True)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"## Add a Trust Layer with Cleanlab\n",
|
||
"\n",
|
||
"Let's add a detection layer to flag untrustworthy RAG responses in real-time. TrustworthyRAG runs Cleanlab's state-of-the-art uncertainty estimator, the [Trustworthy Language Model](https://cleanlab.ai/tlm/), to provide a **trustworthiness score** indicating overall confidence that your RAG's response is *correct*. \n",
|
||
"\n",
|
||
"To diagnose *why* responses are untrustworthy, TrustworthyRAG can run additional evaluations of specific RAG components. Let's see what Evals it runs by default:"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"context_sufficiency\n",
|
||
"response_groundedness\n",
|
||
"response_helpfulness\n",
|
||
"query_ease\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"default_evals = get_default_evals()\n",
|
||
"for eval in default_evals:\n",
|
||
" print(f\"{eval.name}\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"Each Eval returns a score between 0-1 (higher is better) that assesses a different aspect of your RAG system:\n",
|
||
"\n",
|
||
"1. **context_sufficiency**: Evaluates whether the retrieved context contains sufficient information to completely answer the query. A low score indicates that key information is missing from the context (perhaps due to poor retrieval or missing documents).\n",
|
||
"\n",
|
||
"2. **response_groundedness**: Evaluates whether claims/information stated in the response are explicitly supported by the provided context.\n",
|
||
"\n",
|
||
"3. **response_helpfulness**: Evaluates whether the response attempts to answer the user query in a helpful manner.\n",
|
||
"\n",
|
||
"4. **query_ease**: Evaluates whether the user query seems easy for an AI system to properly handle. Complex, vague, tricky, or disgruntled-sounding queries receive lower scores.\n",
|
||
"\n",
|
||
"To run TrustworthyRAG, we need the prompt sent to the LLM, which includes the system message, retrieved chunks, the user's query, and the LLM's response.\n",
|
||
"The event handler defined above provides this prompt.\n",
|
||
"Let's define a helper function to run Cleanlab's detection."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"# Helper function to run real-time Evals\n",
|
||
"def get_eval(query, response, event_handler, evaluator):\n",
|
||
" # Get context used by LLM to generate response\n",
|
||
" context = get_retrieved_context(response)\n",
|
||
" # Get prompt template used to build the prompt\n",
|
||
" pt = event_handler.PROMPT_TEMPLATE\n",
|
||
" # Build prompt\n",
|
||
" full_prompt = pt.format(context_str=context, query_str=query)\n",
|
||
"\n",
|
||
" eval_result = evaluator.score(\n",
|
||
" query=query,\n",
|
||
" context=context,\n",
|
||
" response=response.response,\n",
|
||
" prompt=full_prompt,\n",
|
||
" )\n",
|
||
" # Evaluate the response using TrustworthyRAG\n",
|
||
" print(\"### Evaluation results:\")\n",
|
||
" for metric, value in eval_result.items():\n",
|
||
" print(f\"{metric}: {value['score']}\")\n",
|
||
"\n",
|
||
"\n",
|
||
"# Helper function run end-to-end RAG\n",
|
||
"def get_answer(query, evaluator=trustworthy_rag, event_handler=event_handler):\n",
|
||
" response = query_engine.query(query)\n",
|
||
"\n",
|
||
" print(\n",
|
||
" f\"### Query:\\n{query}\\n\\n### Trimmed Context:\\n{get_retrieved_context(response)[:300]}...\"\n",
|
||
" )\n",
|
||
" print(f\"\\n### Generated response:\\n{response.response}\\n\")\n",
|
||
"\n",
|
||
" get_eval(query, response, event_handler, evaluator)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"### Evaluation results:\n",
|
||
"trustworthiness: 1.0\n",
|
||
"context_sufficiency: 0.9975124377856721\n",
|
||
"response_groundedness: 0.9975124378045552\n",
|
||
"response_helpfulness: 0.9975124367363073\n",
|
||
"query_ease: 0.9975071027792313\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"get_eval(query, response, event_handler, trustworthy_rag)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"**Analysis:** The high `trustworthiness_score` indicates this response is very trustworthy, i.e. non-hallucinated and likely correct. The context that was retrieved here is sufficient to answer this query, as reflected by the high `context_sufficiency` score. The high `query_ease` score indicates this is a straightforward query as well."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"Now let’s run a *challenging* query that **cannot** be answered using the only document in our RAG application's knowledge base."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"### Query:\n",
|
||
"How does the report explain why NVIDIA's Gaming revenue decreased year over year?\n",
|
||
"\n",
|
||
"### Trimmed Context:\n",
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago and up 19% from the previous quarter.\n",
|
||
"\n",
|
||
"- **Quarterly revenue** of $7.19 billion, up 19% from the pre...\n",
|
||
"\n",
|
||
"### Generated response:\n",
|
||
"The report indicates that NVIDIA's Gaming revenue decreased year over year by 38%, which is attributed to a combination of factors, although specific reasons are not detailed. The context highlights that the revenue for the first quarter was $2.24 billion, down from the previous year, while it did show an increase of 22% from the previous quarter. This suggests that while there may have been a seasonal or cyclical recovery, the overall year-over-year decline reflects challenges in the gaming segment during that period.\n",
|
||
"\n",
|
||
"### Evaluation results:\n",
|
||
"trustworthiness: 0.8018049078305449\n",
|
||
"context_sufficiency: 0.26134514055082803\n",
|
||
"response_groundedness: 0.8147481620994604\n",
|
||
"response_helpfulness: 0.28647897539109127\n",
|
||
"query_ease: 0.952132218665045\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"get_answer(\n",
|
||
" \"How does the report explain why NVIDIA's Gaming revenue decreased year over year?\"\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"**Analysis:** The generator LLM avoids conjecture by providing a reliable response, as seen in the high `trustworthiness_score`. The low `context_sufficiency` score reflects that the retrieved context was lacking, and the response doesn’t actually answer the user’s query, as indicated by the low `response_helpfulness`."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"Let’s see how our RAG system responds to another *challenging* question."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"### Query:\n",
|
||
"How much did Nvidia's revenue decrease this quarter vs last quarter, in dollars?\n",
|
||
"\n",
|
||
"### Trimmed Context:\n",
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago and up 19% from the previous quarter.\n",
|
||
"\n",
|
||
"- **Quarterly revenue** of $7.19 billion, up 19% from the pre...\n",
|
||
"\n",
|
||
"### Generated response:\n",
|
||
"NVIDIA's revenue decreased by $1.10 billion this quarter compared to the last quarter.\n",
|
||
"\n",
|
||
"### Evaluation results:\n",
|
||
"trustworthiness: 0.572441384819641\n",
|
||
"context_sufficiency: 0.9974990573223977\n",
|
||
"response_groundedness: 0.006136548076912901\n",
|
||
"response_helpfulness: 0.997512230771839\n",
|
||
"query_ease: 0.8018484929561781\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"get_answer(\n",
|
||
" \"How much did Nvidia's revenue decrease this quarter vs last quarter, in dollars?\"\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"**Analysis**: The generated response incorrectly states that NVIDIA's revenue decreased this quarter, when in fact the referenced report notes a 19% increase quarter-over-quarter. \n",
|
||
"\n",
|
||
"Cleanlab's low trustworthiness score helps us automatically catch this incorrect RAG response in real-time! To root-cause why this response was untrustworthy, we see the `response_groundedness` score is low, which indicates our LLM model is to blame for fabricating this false information. \n",
|
||
"\n",
|
||
"Let's try another one:"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"### Query:\n",
|
||
"If NVIDIA's Data Center segment maintains its Q1 FY2024 quarter-over-quarter growth rate for the next four quarters, what would be its projected annual revenue?\n",
|
||
"\n",
|
||
"### Trimmed Context:\n",
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago and up 19% from the previous quarter.\n",
|
||
"\n",
|
||
"- **Quarterly revenue** of $7.19 billion, up 19% from the pre...\n",
|
||
"\n",
|
||
"### Generated response:\n",
|
||
"If NVIDIA's Data Center segment maintains its quarter-over-quarter growth rate of 18% from Q1 FY2024 for the next four quarters, the projected revenue for the next four quarters can be calculated as follows:\n",
|
||
"\n",
|
||
"1. Q1 FY2024 revenue: $4.28 billion\n",
|
||
"2. Q2 FY2024 projected revenue: $4.28 billion * 1.18 = $5.05 billion\n",
|
||
"3. Q3 FY2024 projected revenue: $5.05 billion * 1.18 = $5.95 billion\n",
|
||
"4. Q4 FY2024 projected revenue: $5.95 billion * 1.18 = $7.02 billion\n",
|
||
"5. Q5 FY2024 projected revenue: $7.02 billion * 1.18 = $8.27 billion\n",
|
||
"\n",
|
||
"Now, summing these revenues for the five quarters (including Q1 FY2024):\n",
|
||
"\n",
|
||
"- Total projected revenue = $4.28 billion + $5.05 billion + $5.95 billion + $7.02 billion + $8.27 billion = $30.57 billion\n",
|
||
"\n",
|
||
"Therefore, the projected annual revenue for the Data Center segment would be approximately $30.57 billion.\n",
|
||
"\n",
|
||
"### Evaluation results:\n",
|
||
"trustworthiness: 0.23124932848015411\n",
|
||
"context_sufficiency: 0.9299227307108295\n",
|
||
"response_groundedness: 0.31247206392894905\n",
|
||
"response_helpfulness: 0.9975055879546202\n",
|
||
"query_ease: 0.7724662723193096\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"get_answer(\n",
|
||
" \"If NVIDIA's Data Center segment maintains its Q1 FY2024 quarter-over-quarter growth rate for the next four quarters, what would be its projected annual revenue?\"\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"**Analysis**: Reviewing the generated response, we find it overstates (sums up the financials of Q1) the projected revenue. Again Cleanlab helps us automatically catch this incorrect response via its low `trustworthiness_score`. Based on the additional Evals, the root cause of this issue again appears to be the LLM model failing to ground its response in the retrieved context."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"### Custom Evals\n",
|
||
"\n",
|
||
"You can also specify custom evaluations to assess specific criteria, and combine them with the default evaluations for comprehensive/tailored assessment of your RAG system.\n",
|
||
"\n",
|
||
"For instance, here's how to create and run a custom eval that checks the conciseness of the generated response."
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"conciseness_eval = Eval(\n",
|
||
" name=\"response_conciseness\",\n",
|
||
" criteria=\"Evaluate whether the Generated response is concise and to the point without unnecessary verbosity or repetition. A good response should be brief but comprehensive, covering all necessary information without extra words or redundant explanations.\",\n",
|
||
" response_identifier=\"Generated Response\",\n",
|
||
")\n",
|
||
"\n",
|
||
"# Combine default evals with a custom eval\n",
|
||
"combined_evals = get_default_evals() + [conciseness_eval]\n",
|
||
"\n",
|
||
"# Initialize TrustworthyRAG with combined evals\n",
|
||
"combined_trustworthy_rag = TrustworthyRAG(evals=combined_evals)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"### Query:\n",
|
||
"What significant transitions did Jensen comment on?\n",
|
||
"\n",
|
||
"### Trimmed Context:\n",
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago and up 19% from the previous quarter.\n",
|
||
"\n",
|
||
"- **Quarterly revenue** of $7.19 billion, up 19% from the pre...\n",
|
||
"\n",
|
||
"### Generated response:\n",
|
||
"Jensen Huang commented on the significant transitions the computer industry is undergoing, particularly in the areas of accelerated computing and generative AI.\n",
|
||
"\n",
|
||
"### Evaluation results:\n",
|
||
"trustworthiness: 0.9810004109697261\n",
|
||
"context_sufficiency: 0.9902170786836257\n",
|
||
"response_groundedness: 0.9975123614036665\n",
|
||
"response_helpfulness: 0.9420916924086002\n",
|
||
"query_ease: 0.5334109647649754\n",
|
||
"response_conciseness: 0.842668665703559\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"get_answer(\n",
|
||
" \"What significant transitions did Jensen comment on?\",\n",
|
||
" evaluator=combined_trustworthy_rag,\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"### Replace your LLM with Cleanlab's\n",
|
||
"\n",
|
||
"Beyond evaluating responses already generated from your LLM, Cleanlab can also generate responses and evaluate them simultaneously (using one of many [supported models](https://help.cleanlab.ai/tlm/api/python/tlm/#class-tlmoptions)). <br />\n",
|
||
"You can do this by calling `trustworthy_rag.generate(query=query, context=context, prompt=full_prompt)` <br />\n",
|
||
"This replaces your own LLM within your RAG system and can be more convenient/accurate/faster.\n",
|
||
"\n",
|
||
"Let's replace our OpenAI LLM to call Cleanlab's endpoint instead:"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"### Query:\n",
|
||
"How much did Nvidia's revenue decrease this quarter vs last quarter, in dollars?\n",
|
||
"\n",
|
||
"### Trimmed Context:\n",
|
||
"# NVIDIA Announces Financial Results for First Quarter Fiscal 2024\n",
|
||
"\n",
|
||
"NVIDIA (NASDAQ: NVDA) today reported revenue for the first quarter ended April 30, 2023, of $7.19 billion, down 13% from a year ago and up 19% from the previous quarter.\n",
|
||
"\n",
|
||
"- **Quarterly revenue** of $7.19 billion, up 19% from the pre\n",
|
||
"\n",
|
||
"### Generated Response:\n",
|
||
"NVIDIA's revenue for the first quarter of fiscal 2024 was $7.19 billion, and for the previous quarter (Q4 FY23), it was $6.05 billion. Therefore, the revenue increased by $1.14 billion from the previous quarter, not decreased. \n",
|
||
"\n",
|
||
"So, the revenue did not decrease this quarter vs last quarter; it actually increased by $1.14 billion.\n",
|
||
"\n",
|
||
"### Evaluation Scores:\n",
|
||
"trustworthiness: 0.6810414232214796\n",
|
||
"context_sufficiency: 0.9974887437375295\n",
|
||
"response_groundedness: 0.9975116791816968\n",
|
||
"response_helpfulness: 0.3293002430120912\n",
|
||
"query_ease: 0.33275910932109172\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"query = \"How much did Nvidia's revenue decrease this quarter vs last quarter, in dollars?\"\n",
|
||
"relevant_chunks = query_engine.retrieve(query)\n",
|
||
"context = get_retrieved_context(relevant_chunks)\n",
|
||
"print(f\"### Query:\\n{query}\\n\\n### Trimmed Context:\\n{context[:300]}\")\n",
|
||
"\n",
|
||
"pt = event_handler.PROMPT_TEMPLATE\n",
|
||
"full_prompt = pt.format(context_str=context, query_str=query)\n",
|
||
"\n",
|
||
"result = trustworthy_rag.generate(\n",
|
||
" query=query, context=context, prompt=full_prompt\n",
|
||
")\n",
|
||
"print(f\"\\n### Generated Response:\\n{result['response']}\\n\")\n",
|
||
"print(\"### Evaluation Scores:\")\n",
|
||
"for metric, value in result.items():\n",
|
||
" if metric != \"response\":\n",
|
||
" print(f\"{metric}: {value['score']}\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"While it remains hard to achieve a RAG application that will accurately answer *any* possible question, you can easily use Cleanlab to deploy a *trustworthy* RAG application which at least flags answers that are likely inaccurate. Learn more about optional configurations you can adjust to improve accuracy/latency in the [Cleanlab documentation](https://help.cleanlab.ai/tlm/)."
|
||
]
|
||
}
|
||
],
|
||
"metadata": {
|
||
"kernelspec": {
|
||
"display_name": "cl",
|
||
"language": "python",
|
||
"name": "python3"
|
||
},
|
||
"language_info": {
|
||
"codemirror_mode": {
|
||
"name": "ipython",
|
||
"version": 3
|
||
},
|
||
"file_extension": ".py",
|
||
"mimetype": "text/x-python",
|
||
"name": "python",
|
||
"nbconvert_exporter": "python",
|
||
"pygments_lexer": "ipython3"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 2
|
||
}
|