Thoughts on using LangChain LCEL with Claude

I got into Natural Language Processing (NLP) and Machine Learning (ML) through Search. And this led me into Generative AI (GenAI), which led me back to Search via Retrieval Augmented Generation (RAG). RAG started out relatively simple -- take a query, generate search results, use search results as context for a Large Language Model (LLM) to generate an abstractive summary of the results. Back when I started on my first "official" GenAI project middle of last year, there were not too many frameworks to support building GenAI components (at least not the prompt based ones), except maybe LangChain, which was just starting out. But prompting as a concept is not too difficult to understand and implement, so thats what we did at the time.

I did have plans to use LangChain in my project once it became more stable, so I started out building my components to be "langchain compliant". But that turned out to be a bad idea as LangChain continued its exponential (and from the outside at least, somewhat haphazard) growth and showed no signs of stabilizing. At one point, LangChain users were advised to make pip install -U langchain part of their daily morning routine! So anyway, we ended up building up our GenAI application by hooking up third party components with our own (non-framework) code, using Anthropic's Claude-v2 as our LLM, ElasticSearch as our lexical / vector document store and PostgreSQL as our conversational buffer.

While I continue to believe that the decision to go with our own code made more sense than trying to jump on the LangChain (or Semantic Kernel, or Haystack, or some other) train, I do regret it in some ways. A collateral benefit for people who adopted and stuck with LangChain were the ready-to-use implementations of cutting-edge RAG and GenAI techniques that the community implemented at almost the same pace as they were being proposed in academic papers. For the subset of these people that were even slightly curious about how these implementations worked, this offered a ringside view into the latest advances in the field and a chance to stay current with it, with minimal effort.

So anyway, in an attempt to replicate this benefit for myself (going forward at least), I decided to learn LangChain by doing a small side project. Earlier I needed to learn to use Snowflake for something else and had their free O'Reilly book on disk, so I converted it to text, chunked it, and put it into a Chroma vector store. I then tried to implement examples from the DeepLearning.AI courses LangChain: Chat with your Data and LangChain for LLM Application Development. The big difference is that the course examples use OpenAI's GPT-3 as their LLM whereas I use Claude-2 on AWS Bedrock in mine. In this post, I share the issues I faced and my solutions, hopefully this can help guide others in similar situations.

Couple of observations here. First, the granularity of GenAI components is necessarily larger than traditional software components, and this means application details that the developer of the component was working on can leak into the component itself (mostly through the prompt). To a user of the component, this can manifest as subtle bugs. Fortunately, LangChain developers seem to have also noticed this and have come up with the LangChain Expression Language (LCEL), a small set of reusable components that can be composed to create chains from the ground up. They have also marked a large number of Chains as Legacy Chains (to be converted to LCEL chains in the future).

Second, most of the components (or chains, since that is LangChain's central abstraction) are developed against OpenAI GPT-3 (or its chat version GPT-3.5 Turbo) whose strengths and weaknesses may be different from those of your LLM. For example, OpenAI is very good at generating JSON output, whereas Claude is better at generating XML. I have also seen that Claude can terminate XML / JSON output mid-output unless forced to complete using stop_sequences. Yhis doesn't seem to be a problem GPT-3 users have observed -- when I mentioned this problem and the fix, I drew a blank on both counts.

To address the first issue, my general approach in trying to re-implement these examples has been to use LCEL to build my chains from scratch. I attempt to leverage the expertise available in LangChain by looking in the code or running the existing LangChain chain with langchain.debug set to True. Doing this helps me see the prompt being used and the flow, which I can use to adapt the prompt and flow for my LCEL chain. To address the second issue, I play to Claude's strengths by specifying XML output format in my prompts and parsing them as Pydantic objects for data transfer across chains.

The example application I will use to illustrate these techniques here is derived from the Evaluation lesson from the LangChain for LLM Application Development course, and is illustrated in the diagram below. The application takes a chunk of text as input, and uses the Question Generation chain to generate multiple question-answer pairs from it. The questions and the original content are fed into the Question Answering chain, which uses the question to generate additional context from a vector retriever, and uses all three to generate an answer. The answer generated from the Question Generation chain and the answer generated from the Question Answering chain are fed into a Question Generation Evaluation chain, where the LLM grades one against the other, and generates an aggregate score for the questions generated from the chunk.

Each chain in this pipeline is actually quite simple, they take one or more inputs and generates a block of XML. All the chains are structured as follows:

1 2 3

from langchain_core.output_parsers import StrOutputParser chain = prompt | model | StrOutputParser()

And all our prompts follow the same general format. Here is the prompt for the Evaluation chain (the third one) which I adapted from the QAEvalChain used in the lesson notebook. Developing from scratch using LCEL gives me the chance to use Claude's Human / Assistant format (see LangChain Guidelines for Anthropic) rather than depend on the generic prompt that happens to work well for GPT-3.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34

Human: You are a teacher grading a quiz. You are given a question, the context the question is about, and the student's answer. QUESTION: {question} CONTEXT: {context} STUDENT ANSWER: {predicted_answer} TRUE ANSWER: {generated_answer} You are to score the student's answer as either CORRECT or INCORRECT, based on the context. Write out in a step by step manner your reasoning to be sure that your conclusion is correct. Avoid simply stating the correct answer at the outset. Please provide your response in the following format: <result> <qa_eval> <question>the question here</question> <student_answer>the student's answer here</student_answer> <true_answer>the true answer here</true_answer> <explanation>step by step reasoning here</explanation> <grade>CORRECT or INCORRECT here</grade> </qa_eval> </result> Grade the student answers based ONLY on their factual accuracy. Ignore differences in punctuation and phrasing between the student answer and true answer. It is OK if the student answer contains more information than the true answer, as long as it does not contain any conflicting statements. Assistant:

In addition, I specify the formatting instructions explicitly in the prompt instead of using the canned ones from XMLOutputParser or PydanticOutputParser via get_formatting_instructions(), which are comparatively quite generic and sub-optimal. By convention, the outermost tag in my format is always <result>...</result>. The qa_eval tag inside result has a corresponding Pydantic class analog declared in the code as follows:

1 2 3 4 5 6 7 8 9 10 11 12

from pydantic import BaseModel, Field class QAEval(BaseModel): question: str = Field(alias="question", description="question text") student_answer: str = Field(alias="student_answer", description="answer predicted by QA chain") true_answer: str = Field(alias="true_answer", description="answer generated by QG chain") explanation: str = Field(alias="explanation", description="chain of thought for grading") grade: str = Field(alias="grade", description="LLM grade CORRECT or INCORRECT")

After the StrOutputParser extracts the LLM output into a string, it is first passed through a regular expression to remove any content outside the <result>...</result>, then convert it into the QAEval Pydantic object using the following code. This allows us to keep object manipulation between chains independent of the output format, as well as negate any need for format specific parsing.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

import re import xmltodict from pydantic import Field from pydantic.generics import GenericModel from typing import Generic, List, Tuple, TypeVar T = TypeVar("T") class Result(GenericModel, Generic[T]): value: T = Field(alias="result") def parse_response(response): response = response.strip() start_tag, end_tag = "<result>", "</result>" is_valid = response.startswith(start_tag) and response.endswith(end_tag) if not is_valid: pattern = f"(?:{start_tag})(.*)(?:{end_tag})" p = re.compile(pattern, re.DOTALL) m = p.search(response) if m is not None: response = start_tag + m.group(1) + end_tag resp_dict = xmltodict.parse(response) result = Result(**resp_dict) return result # example call response = chain.invoke( "question": "the question", "context": "the context", "predicted_answer": "the predicted answer", "generated_answer": "the generated answer" }) result = parse_response(response) qa_eval = result.value["qa_eval"]

One downside to this approach is that it uses the current version of the Pydantic toolkit (v2) whereas LangChain still uses Pydantic V1 internally, as descibed in LangChain's Pydantic compatibility page. This is why this conversion needs to be outside LangChain and in the application code. Ideally, I would like this to be part of a subclass of PydanticOutputParser where the formatting_instructions could be generated from the class definition as a nice side effect, but that would mean more work than I am prepared to do at this point :-). Meanwhile, this seems like a decent compromise.

Thats all I had for today. Thank you for staying with me so far, and hope you found this useful!