Tip 2: Use structured outputs
Utilizing structured outputs means forcing the LLM to output legitimate JSON or YAML textual content. This can permit you to cut back the ineffective ramblings and get “straight-to-the-point” solutions about what you want from the LLM. It additionally will assist with the subsequent suggestions because it makes the LLM responses simpler to confirm.
Right here is how you are able to do this with Gemini’s API:
import jsonimport google.generativeai as genai
from pydantic import BaseModel, Subject
from document_ai_agents.schema_utils import prepare_schema_for_gemini
class Reply(BaseModel):
reply: str = Subject(..., description="Your Reply.")
mannequin = genai.GenerativeModel("gemini-1.5-flash-002")
answer_schema = prepare_schema_for_gemini(Reply)
query = "Record all of the explanation why LLM hallucinate"
context = (
"LLM hallucination refers back to the phenomenon the place massive language fashions generate plausible-sounding however"
" factually incorrect or nonsensical info. This may happen as a result of numerous components, together with biases"
" within the coaching information, the inherent limitations of the mannequin's understanding of the actual world, and the "
"mannequin's tendency to prioritize fluency and coherence over accuracy."
)
messages = (
[context]
+ [
f"Answer this question: {question}",
]
+ [
f"Use this schema for your answer: {answer_schema}",
]
)
response = mannequin.generate_content(
messages,
generation_config={
"response_mime_type": "software/json",
"response_schema": answer_schema,
"temperature": 0.0,
},
)
response = Reply(**json.masses(response.textual content))
print(f"{response.reply=}")
The place “prepare_schema_for_gemini” is a utility perform that prepares the schema to match Gemini’s bizarre necessities. You’ll find its definition right here: code.
This code defines a Pydantic schema and sends this schema as a part of the question within the discipline “response_schema”. This forces the LLM to comply with this schema in its response and makes it simpler to parse its output.
Tip 3: Use chain of ideas and higher prompting
Generally, giving the LLM the area to work out its response, earlier than committing to a ultimate reply, will help produce higher high quality responses. This method is named Chain-of-thoughts and is broadly used as it’s efficient and really simple to implement.
We will additionally explicitly ask the LLM to reply with “N/A” if it might probably’t discover sufficient context to supply a high quality response. This can give it a straightforward method out as an alternative of making an attempt to answer questions it has no reply to.
For instance, lets look into this easy query and context:
Context
Thomas Jefferson (April 13 [O.S. April 2], 1743 — July 4, 1826) was an American statesman, planter, diplomat, lawyer, architect, thinker, and Founding Father who served because the third president of the US from 1801 to 1809.[6] He was the first writer of the Declaration of Independence. Following the American Revolutionary Battle and earlier than turning into president in 1801, Jefferson was the nation’s first U.S. secretary of state underneath George Washington after which the nation’s second vp underneath John Adams. Jefferson was a number one proponent of democracy, republicanism, and pure rights, and he produced formative paperwork and selections on the state, nationwide, and worldwide ranges. (Supply: Wikipedia)
Query
What yr did davis jefferson die?
A naive method yields:
Response
reply=’1826′
Which is clearly false as Jefferson Davis shouldn’t be even talked about within the context in any respect. It was Thomas Jefferson that died in 1826.
If we modify the schema of the response to make use of chain-of-thoughts to:
class AnswerChainOfThoughts(BaseModel):
rationale: str = Subject(
...,
description="Justification of your reply.",
)
reply: str = Subject(
..., description="Your Reply. Reply with 'N/A' if reply shouldn't be discovered"
)
We’re additionally including extra particulars about what we count on as output when the query shouldn’t be answerable utilizing the context “Reply with ‘N/A’ if reply shouldn’t be discovered”
With this new method, we get the next rationale (keep in mind, chain-of-thought):
The offered textual content discusses Thomas Jefferson, not Jefferson Davis. No details about the dying of Jefferson Davis is included.
And the ultimate reply:
reply=’N/A’
Nice ! However can we use a extra common method to hallucination detection?
We will, with Brokers!
Tip 4: Use an Agentic method
We’ll construct a easy agent that implements a three-step course of:
- Step one is to incorporate the context and ask the query to the LLM with a purpose to get the primary candidate response and the related context that it had used for its reply.
- The second step is to reformulate the query and the primary candidate response as a declarative assertion.
- The third step is to ask the LLM to confirm whether or not or not the related context entails the candidate response. It’s referred to as “Self-verification”: https://arxiv.org/pdf/2212.09561
As a way to implement this, we outline three nodes in LangGraph. The primary node will ask the query whereas together with the context, the second node will reformulate it utilizing the LLM and the third node will test the entailment of the assertion in relation to the enter context.
The primary node will be outlined as follows:
def answer_question(self, state: DocumentQAState):
logger.data(f"Responding to query '{state.query}'")
assert (
state.pages_as_base64_jpeg_images or state.pages_as_text
), "Enter textual content or pictures"
messages = (
[
{"mime_type": "image/jpeg", "data": base64_jpeg}
for base64_jpeg in state.pages_as_base64_jpeg_images
]
+ state.pages_as_text
+ [
f"Answer this question: {state.question}",
]
+ [
f"Use this schema for your answer: {self.answer_cot_schema}",
]
)response = self.mannequin.generate_content(
messages,
generation_config={
"response_mime_type": "software/json",
"response_schema": self.answer_cot_schema,
"temperature": 0.0,
},
)
answer_cot = AnswerChainOfThoughts(**json.masses(response.textual content))
return {"answer_cot": answer_cot}
And the second as:
def reformulate_answer(self, state: DocumentQAState):
logger.data("Reformulating reply")
if state.answer_cot.reply == "N/A":
returnmessages = [
{
"role": "user",
"parts": [
{
"text": "Reformulate this question and its answer as a single assertion."
},
{"text": f"Question: {state.question}"},
{"text": f"Answer: {state.answer_cot.answer}"},
]
+ [
{
"text": f"Use this schema for your answer: {self.declarative_answer_schema}"
}
],
}
]
response = self.mannequin.generate_content(
messages,
generation_config={
"response_mime_type": "software/json",
"response_schema": self.declarative_answer_schema,
"temperature": 0.0,
},
)
answer_reformulation = AnswerReformulation(**json.masses(response.textual content))
return {"answer_reformulation": answer_reformulation}
The third one as:
def verify_answer(self, state: DocumentQAState):
logger.data(f"Verifying reply '{state.answer_cot.reply}'")
if state.answer_cot.reply == "N/A":
return
messages = [
{
"role": "user",
"parts": [
{
"text": "Analyse the following context and the assertion and decide whether the context "
"entails the assertion or not."
},
{"text": f"Context: {state.answer_cot.relevant_context}"},
{
"text": f"Assertion: {state.answer_reformulation.declarative_answer}"
},
{
"text": f"Use this schema for your answer: {self.verification_cot_schema}. Be Factual."
},
],
}
]response = self.mannequin.generate_content(
messages,
generation_config={
"response_mime_type": "software/json",
"response_schema": self.verification_cot_schema,
"temperature": 0.0,
},
)
verification_cot = VerificationChainOfThoughts(**json.masses(response.textual content))
return {"verification_cot": verification_cot}
Full code in https://github.com/CVxTz/document_ai_agents
Discover how every node makes use of its personal schema for structured output and its personal immediate. That is attainable as a result of flexibility of each Gemini’s API and LangGraph.
Lets work by this code utilizing the identical instance as above ➡️
(Notice: we aren’t utilizing chain-of-thought on the primary immediate in order that the verification will get triggered for our checks.)
Context
Thomas Jefferson (April 13 [O.S. April 2], 1743 — July 4, 1826) was an American statesman, planter, diplomat, lawyer, architect, thinker, and Founding Father who served because the third president of the US from 1801 to 1809.[6] He was the first writer of the Declaration of Independence. Following the American Revolutionary Battle and earlier than turning into president in 1801, Jefferson was the nation’s first U.S. secretary of state underneath George Washington after which the nation’s second vp underneath John Adams. Jefferson was a number one proponent of democracy, republicanism, and pure rights, and he produced formative paperwork and selections on the state, nationwide, and worldwide ranges. (Supply: Wikipedia)
Query
What yr did davis jefferson die?
First node outcome (First reply):
relevant_context=’Thomas Jefferson (April 13 [O.S. April 2], 1743 — July 4, 1826) was an American statesman, planter, diplomat, lawyer, architect, thinker, and Founding Father who served because the third president of the US from 1801 to 1809.’
reply=’1826′
Second node outcome (Reply Reformulation):
declarative_answer=’Davis Jefferson died in 1826′
Third node outcome (Verification):
rationale=’The context states that Thomas Jefferson died in 1826. The assertion states that Davis Jefferson died in 1826. The context doesn’t point out Davis Jefferson, solely Thomas Jefferson.’
entailment=’No’
So the verification step rejected (No entailment between the 2) the preliminary reply. We will now keep away from returning a hallucination to the person.
Bonus Tip : Use stronger fashions
This tip shouldn’t be all the time simple to use as a result of funds or latency limitations however you need to know that stronger LLMs are much less susceptible to hallucination. So, if attainable, go for a extra highly effective LLM in your most delicate use circumstances. You may test a benchmark of hallucinations right here: https://github.com/vectara/hallucination-leaderboard. We will see that the highest fashions on this benchmark (least hallucinations) additionally ranks on the high of standard NLP chief boards.
On this tutorial, we explored methods to enhance the reliability of LLM outputs by decreasing the hallucination charge. The primary suggestions embrace cautious formatting and prompting to information LLM calls and utilizing a workflow primarily based method the place Brokers are designed to confirm their very own solutions.
This includes a number of steps:
- Retrieving the precise context components utilized by the LLM to generate the reply.
- Reformulating the reply for simpler verification (In declarative kind).
- Instructing the LLM to test for consistency between the context and the reformulated reply.
Whereas all the following pointers can considerably enhance accuracy, you need to keep in mind that no technique is foolproof. There’s all the time a danger of rejecting legitimate solutions if the LLM is overly conservative throughout verification or lacking actual hallucination circumstances. Subsequently, rigorous analysis of your particular LLM workflows remains to be important.
Full code in https://github.com/CVxTz/document_ai_agents