Agentic AI from First Rules: Reflection

says that “any sufficiently superior expertise is indistinguishable from magic”. That’s precisely how lots of right now’s AI frameworks really feel. Instruments like GitHub Copilot, Claude Desktop, OpenAI Operator, and Perplexity Comet are automating on a regular basis duties that will’ve appeared not possible to automate simply 5 years in the past. What’s much more outstanding is that with just some traces of code, we are able to construct our personal subtle AI instruments: ones that search via information, browse the online, click on hyperlinks, and even make purchases. It actually does really feel like magic.

Although I genuinely imagine in knowledge wizards, I don’t imagine in magic. I discover it thrilling (and sometimes useful) to know how issues are literally constructed and what’s taking place below the hood. That’s why I’ve determined to share a collection of posts on agentic AI design ideas that’ll make it easier to perceive how all these magical instruments truly work.

To achieve a deep understanding, we’ll construct a multi-AI agent system from scratch. We’ll keep away from utilizing frameworks like CrewAI or smolagents and as a substitute work immediately with the muse mannequin API. Alongside the best way, we’ll discover the elemental agentic design patterns: reflection, device use, planning, and multi-agent setups. Then, we’ll mix all this information to construct a multi-AI agent system that may reply advanced data-related questions.

As Richard Feynman put it, “What I can’t create, I don’t perceive.” So let’s begin constructing! On this article, we’ll deal with the reflection design sample. However first, let’s determine what precisely reflection is.

What reflection is

Let’s mirror on how we (people) often work on duties. Think about I must share the outcomes of a current characteristic launch with my PM. I’ll possible put collectively a fast draft after which learn it a few times from starting to finish, guaranteeing that each one elements are constant, there’s sufficient info, and there are not any typos.

Or let’s take one other instance: writing a SQL question. I’ll both write it step-by-step, checking the intermediate outcomes alongside the best way, or (if it’s easy sufficient) I’ll draft it suddenly, execute it, have a look at the outcome (checking for errors or whether or not the outcome matches my expectations), after which tweak the question based mostly on that suggestions. I’d rerun it, test the outcome, and iterate till it’s proper.

So we hardly ever write lengthy texts from prime to backside in a single go. We often circle again, evaluate, and tweak as we go. These suggestions loops are what assist us enhance the standard of our work.

LLMs use a special strategy. When you ask an LLM a query, by default, it’ll generate a solution token by token, and the LLM gained’t be capable to evaluate its outcome and repair any points. However in an agentic AI setup, we are able to create suggestions loops for LLMs too, both by asking the LLM to evaluate and enhance its personal reply or by sharing exterior suggestions with it (just like the outcomes of a SQL execution). And that’s the entire level of reflection. It sounds fairly simple, however it could yield considerably higher outcomes.

There’s a considerable physique of analysis displaying the advantages of reflection:

Picture from “Self-Refine: Iterative Refinement with Self-Suggestions,” Madaan et al.

In “Reflexion: Language Brokers with Verbal Reinforcement Studying” Shinn et al. (2023), the authors achieved a 91% move@1 accuracy on the HumanEval coding benchmark, surpassing the earlier state-of-the-art GPT-4, which scored simply 80%. In addition they discovered that Reflexion considerably outperforms all baseline approaches on the HotPotQA benchmark (a Wikipedia-based Q&A dataset that challenges brokers to parse content material and cause over a number of supporting paperwork).

Picture from “Reflexion: Language Brokers with Verbal Reinforcement Studying,” Shinn et al.

Reflection is very impactful in agentic techniques as a result of it may be used to course-correct at many steps of the method:

When a consumer asks a query, the LLM can use reflection to guage whether or not the request is possible.
When the LLM places collectively an preliminary plan, it could use reflection to double-check whether or not the plan is sensible and may also help obtain the purpose.
After every execution step or device name, the agent can consider whether or not it’s on monitor and whether or not it’s value adjusting the plan.
When the plan is absolutely executed, the agent can mirror to see whether or not it has truly achieved the purpose and solved the duty.

It’s clear that reflection can considerably enhance accuracy. Nonetheless, there are trade-offs value discussing. Reflection may require a number of further calls to the LLM and probably different techniques, which might result in elevated latency and prices. So in enterprise circumstances, it’s value contemplating whether or not the standard enhancements justify the bills and delays within the consumer circulation.

Reflection in frameworks

Since there’s little doubt that reflection brings worth to AI brokers, it’s broadly utilized in fashionable frameworks. Let’s have a look at some examples.

The concept of reflection was first proposed within the paper “ReAct: Synergizing Reasoning and Appearing in Language Fashions” by Yao et al. (2022). ReAct is a framework that mixes interleaving levels of Reasoning (reflection via specific thought traces) and Appearing (task-relevant actions in an surroundings). On this framework, reasoning guides the selection of actions, and actions produce new observations that inform additional reasoning. The reasoning stage itself is a mixture of reflection and planning.

This framework grew to become fairly fashionable, so there at the moment are a number of off-the-shelf implementations, reminiscent of:

The DSPy framework by Databricks has a ReAct class,
In LangGraph, you should use the create_react_agent perform,
Code brokers within the smolagents library by HuggingFace are additionally based mostly on the ReAct structure.

Reflection from scratch

Now that we’ve realized the speculation and explored current implementations, it’s time to get our fingers soiled and construct one thing ourselves. Within the ReAct strategy, brokers use reflection at every step, combining planning with reflection. Nonetheless, to know the affect of reflection extra clearly, we’ll have a look at it in isolation.

For instance, we’ll use text-to-SQL: we’ll give an LLM a query and count on it to return a sound SQL question. We’ll be working with a flight delay dataset and the ClickHouse SQL dialect.

We’ll begin through the use of direct era with none reflection as our baseline. Then, we’ll attempt utilizing reflection by asking the mannequin to critique and enhance the SQL, or by offering it with further suggestions. After that, we’ll measure the standard of our solutions to see whether or not reflection truly results in higher outcomes.

Direct era

We’ll start with probably the most simple strategy, direct era, the place we ask the LLM to generate SQL that solutions a consumer question.

pip set up anthropic

We have to specify the API Key for the Anthropic API.

import os
os.environ['ANTHROPIC_API_KEY'] = config['ANTHROPIC_API_KEY']

The subsequent step is to initialise the shopper, and we’re all set.

import anthropic
shopper = anthropic.Anthropic()

Now we are able to use this shopper to ship messages to the LLM. Let’s put collectively a perform to generate SQL based mostly on a consumer question. I’ve specified the system immediate with primary directions and detailed details about the info schema. I’ve additionally created a perform to ship the system immediate and consumer question to the LLM.

base_sql_system_prompt = '''
You're a senior SQL developer and your activity is to assist generate a SQL question based mostly on consumer necessities. 
You might be working with ClickHouse database. Specify the format (Tab Separated With Names) within the SQL question output to make sure that column names are included within the output.
Don't use rely(*) in your queries since it is a dangerous observe with columnar databases, favor utilizing rely().
Be certain that the question is syntactically appropriate and optimized for efficiency, considering ClickHouse particular options (i.e. that ClickHouse is a columnar database and helps capabilities like ARRAY JOIN, SAMPLE, and so on.).
Return solely the SQL question with none further explanations or feedback.

You'll be working with flight_data desk which has the next schema:

Column Identify | Information Kind | Null % | Instance Worth | Description
--- | --- | --- | --- | ---
yr | Int64 | 0.0 | 2024 | Yr of flight
month | Int64 | 0.0 | 1 | Month of flight (1–12)
day_of_month | Int64 | 0.0 | 1 | Day of the month
day_of_week | Int64 | 0.0 | 1 | Day of week (1=Monday … 7=Sunday)
fl_date | datetime64[ns] | 0.0 | 2024-01-01 00:00:00 | Flight date (YYYY-MM-DD)
op_unique_carrier | object | 0.0 | 9E | Distinctive provider code
op_carrier_fl_num | float64 | 0.0 | 4814.0 | Flight quantity for reporting airline
origin | object | 0.0 | JFK | Origin airport code
origin_city_name | object | 0.0 | "New York, NY" | Origin metropolis identify
origin_state_nm | object | 0.0 | New York | Origin state identify
dest | object | 0.0 | DTW | Vacation spot airport code
dest_city_name | object | 0.0 | "Detroit, MI" | Vacation spot metropolis identify
dest_state_nm | object | 0.0 | Michigan | Vacation spot state identify
crs_dep_time | Int64 | 0.0 | 1252 | Scheduled departure time (native, hhmm)
dep_time | float64 | 1.31 | 1247.0 | Precise departure time (native, hhmm)
dep_delay | float64 | 1.31 | -5.0 | Departure delay in minutes (unfavourable if early)
taxi_out | float64 | 1.35 | 31.0 | Taxi out time in minutes
wheels_off | float64 | 1.35 | 1318.0 | Wheels-off time (native, hhmm)
wheels_on | float64 | 1.38 | 1442.0 | Wheels-on time (native, hhmm)
taxi_in | float64 | 1.38 | 7.0 | Taxi in time in minutes
crs_arr_time | Int64 | 0.0 | 1508 | Scheduled arrival time (native, hhmm)
arr_time | float64 | 1.38 | 1449.0 | Precise arrival time (native, hhmm)
arr_delay | float64 | 1.61 | -19.0 | Arrival delay in minutes (unfavourable if early)
cancelled | int64 | 0.0 | 0 | Cancelled flight indicator (0=No, 1=Sure)
cancellation_code | object | 98.64 | B | Purpose for cancellation (if cancelled)
diverted | int64 | 0.0 | 0 | Diverted flight indicator (0=No, 1=Sure)
crs_elapsed_time | float64 | 0.0 | 136.0 | Scheduled elapsed time in minutes
actual_elapsed_time | float64 | 1.61 | 122.0 | Precise elapsed time in minutes
air_time | float64 | 1.61 | 84.0 | Flight time in minutes
distance | float64 | 0.0 | 509.0 | Distance between origin and vacation spot (miles)
carrier_delay | int64 | 0.0 | 0 | Service-related delay in minutes
weather_delay | int64 | 0.0 | 0 | Climate-related delay in minutes
nas_delay | int64 | 0.0 | 0 | Nationwide Air System delay in minutes
security_delay | int64 | 0.0 | 0 | Safety delay in minutes
late_aircraft_delay | int64 | 0.0 | 0 | Late plane delay in minutes
'''

def generate_direct_sql(rec):
  # making an LLM name
  message = shopper.messages.create(
    mannequin = "claude-3-5-haiku-latest",
    # I selected smaller mannequin in order that it is simpler for us to see the affect 
    max_tokens = 8192,
    system=base_sql_system_prompt,
    messages = [
        {'role': 'user', 'content': rec['question']}
    ]
  )

  sql  = message.content material[0].textual content
  
  # cleansing the output
  if sql.endswith('```'):
    sql = sql[:-3]
  if sql.startswith('```sql'):
    sql = sql[6:]
  return sql

That’s it. Now let’s take a look at our text-to-SQL answer. I’ve created a small analysis set of 20 question-and-answer pairs that we are able to use to test whether or not our system is working nicely. Right here’s one instance:

{
'query': 'What was the very best velocity in mph?',
'reply': '''
    choose max(distance / (air_time / 60)) as max_speed 
    from flight_data 
    the place air_time > 0 
    format TabSeparatedWithNames'''
}

Let’s use our text-to-SQL perform to generate SQL for all consumer queries within the take a look at set.

# load analysis set
with open('./knowledge/flight_data_qa_pairs.json', 'r') as f:
    qa_pairs = json.load(f)
qa_pairs_df = pd.DataFrame(qa_pairs)

tmp = []
# executing LLM for every query in our eval set
for rec in tqdm.tqdm(qa_pairs_df.to_dict('data')):
    llm_sql = generate_direct_sql(rec)
    tmp.append(
        {
            'id': rec['id'],
            'llm_direct_sql': llm_sql
        }
    )

llm_direct_df = pd.DataFrame(tmp)
direct_result_df = qa_pairs_df.merge(llm_direct_df, on = 'id')

Now we’ve got our solutions, and the subsequent step is to measure the standard.

Measuring high quality

Sadly, there’s no single appropriate reply on this state of affairs, so we are able to’t simply evaluate the SQL generated by the LLM to a reference reply. We have to provide you with a approach to measure high quality.

There are some features of high quality that we are able to test with goal standards, however to test whether or not the LLM returned the fitting reply, we’ll want to make use of an LLM. So I’ll use a mixture of approaches:

First, we’ll use goal standards to test whether or not the proper format was specified within the SQL (we instructed the LLM to make use of TabSeparatedWithNames).
Second, we are able to execute the generated question and see whether or not ClickHouse returns an execution error.
Lastly, we are able to create an LLM decide that compares the output from the generated question to our reference reply and checks whether or not they differ.

Let’s begin by executing the SQL. It’s value noting that our get_clickhouse_data perform doesn’t throw an exception. As a substitute, it returns textual content explaining the error, which could be dealt with by the LLM later.

CH_HOST = 'http://localhost:8123' # default tackle 
import requests
import pandas as pd
import tqdm

# perform to execute SQL question
def get_clickhouse_data(question, host = CH_HOST, connection_timeout = 1500):
  r = requests.publish(host, params = {'question': question}, 
    timeout = connection_timeout)
  if r.status_code == 200:
      return r.textual content
  else: 
      return 'Database returned the next error:n' + r.textual content

# getting the outcomes of SQL execution
direct_result_df['llm_direct_output'] = direct_result_df['llm_direct_sql'].apply(get_clickhouse_data)
direct_result_df['answer_output'] = direct_result_df['answer'].apply(get_clickhouse_data)

The subsequent step is to create an LLM decide. For this, I’m utilizing a series‑of‑thought strategy that prompts the LLM to offer its reasoning earlier than giving the ultimate reply. This provides the mannequin time to assume via the issue, which improves response high quality.

llm_judge_system_prompt = '''
You're a senior analyst and your activity is to match two SQL question outcomes and decide if they're equal. 
Focus solely on the info returned by the queries, ignoring any formatting variations. 
Take note of the preliminary consumer question and data wanted to reply it. For instance, if consumer requested for the common distance, and each queries return the identical common worth however in certainly one of them there's additionally a rely of data, it is best to contemplate them equal, since each present the identical requested info.

Reply with a JSON of the next construction:
{
  'reasoning': '', 
  'equivalence': 
}
Be certain that ONLY JSON is within the output. 

You'll be working with flight_data desk which has the next schema:
Column Identify | Information Kind | Null % | Instance Worth | Description
--- | --- | --- | --- | ---
yr | Int64 | 0.0 | 2024 | Yr of flight
month | Int64 | 0.0 | 1 | Month of flight (1–12)
day_of_month | Int64 | 0.0 | 1 | Day of the month
day_of_week | Int64 | 0.0 | 1 | Day of week (1=Monday … 7=Sunday)
fl_date | datetime64[ns] | 0.0 | 2024-01-01 00:00:00 | Flight date (YYYY-MM-DD)
op_unique_carrier | object | 0.0 | 9E | Distinctive provider code
op_carrier_fl_num | float64 | 0.0 | 4814.0 | Flight quantity for reporting airline
origin | object | 0.0 | JFK | Origin airport code
origin_city_name | object | 0.0 | "New York, NY" | Origin metropolis identify
origin_state_nm | object | 0.0 | New York | Origin state identify
dest | object | 0.0 | DTW | Vacation spot airport code
dest_city_name | object | 0.0 | "Detroit, MI" | Vacation spot metropolis identify
dest_state_nm | object | 0.0 | Michigan | Vacation spot state identify
crs_dep_time | Int64 | 0.0 | 1252 | Scheduled departure time (native, hhmm)
dep_time | float64 | 1.31 | 1247.0 | Precise departure time (native, hhmm)
dep_delay | float64 | 1.31 | -5.0 | Departure delay in minutes (unfavourable if early)
taxi_out | float64 | 1.35 | 31.0 | Taxi out time in minutes
wheels_off | float64 | 1.35 | 1318.0 | Wheels-off time (native, hhmm)
wheels_on | float64 | 1.38 | 1442.0 | Wheels-on time (native, hhmm)
taxi_in | float64 | 1.38 | 7.0 | Taxi in time in minutes
crs_arr_time | Int64 | 0.0 | 1508 | Scheduled arrival time (native, hhmm)
arr_time | float64 | 1.38 | 1449.0 | Precise arrival time (native, hhmm)
arr_delay | float64 | 1.61 | -19.0 | Arrival delay in minutes (unfavourable if early)
cancelled | int64 | 0.0 | 0 | Cancelled flight indicator (0=No, 1=Sure)
cancellation_code | object | 98.64 | B | Purpose for cancellation (if cancelled)
diverted | int64 | 0.0 | 0 | Diverted flight indicator (0=No, 1=Sure)
crs_elapsed_time | float64 | 0.0 | 136.0 | Scheduled elapsed time in minutes
actual_elapsed_time | float64 | 1.61 | 122.0 | Precise elapsed time in minutes
air_time | float64 | 1.61 | 84.0 | Flight time in minutes
distance | float64 | 0.0 | 509.0 | Distance between origin and vacation spot (miles)
carrier_delay | int64 | 0.0 | 0 | Service-related delay in minutes
weather_delay | int64 | 0.0 | 0 | Climate-related delay in minutes
nas_delay | int64 | 0.0 | 0 | Nationwide Air System delay in minutes
security_delay | int64 | 0.0 | 0 | Safety delay in minutes
late_aircraft_delay | int64 | 0.0 | 0 | Late plane delay in minutes
'''

llm_judge_user_prompt_template = '''
Right here is the preliminary consumer question:
{user_query}

Right here is the SQL question generated by the primary analyst: 
SQL: 
{sql1} 

Database output: 
{result1}

Right here is the SQL question generated by the second analyst:
SQL:
{sql2}

Database output:
{result2}
'''

def llm_judge(rec, field_to_check):
  # assemble the consumer immediate 
  user_prompt = llm_judge_user_prompt_template.format(
    user_query = rec['question'],
    sql1 = rec['answer'],
    result1 = rec['answer_output'],
    sql2 = rec[field_to_check + '_sql'],
    result2 = rec[field_to_check + '_output']
  )
  
  # make an LLM name
  message = shopper.messages.create(
      mannequin = "claude-sonnet-4-5",
      max_tokens = 8192,
      temperature = 0.1,
      system = llm_judge_system_prompt,
      messages=[
          {'role': 'user', 'content': user_prompt}
      ]
  )
  knowledge = message.content material[0].textual content
  
  # Strip markdown code blocks
  knowledge = knowledge.strip()
  if knowledge.startswith('```json'):
      knowledge = knowledge[7:]
  elif knowledge.startswith('```'):
      knowledge = knowledge[3:]
  if knowledge.endswith('```'):
      knowledge = knowledge[:-3]
  
  knowledge = knowledge.strip()
  return json.masses(knowledge)

Now, let’s run the LLM decide to get the outcomes.

tmp = []

for rec in tqdm.tqdm(direct_result_df.to_dict('data')):
  attempt:
    judgment = llm_judge(rec, 'llm_direct')
  besides Exception as e:
    print(f"Error processing report {rec['id']}: {e}")
    proceed
  tmp.append(
    {
      'id': rec['id'],
      'llm_judge_reasoning': judgment['reasoning'],
      'llm_judge_equivalence': judgment['equivalence']
    }
  )

judge_df = pd.DataFrame(tmp)
direct_result_df = direct_result_df.merge(judge_df, on = 'id')

Let’s have a look at one instance to see how the LLM decide works.

# consumer question 
In 2024, what share of time all airplanes spent within the air?

# appropriate reply 
choose (sum(air_time) / sum(actual_elapsed_time)) * 100 as percentage_in_air 
the place yr = 2024
from flight_data 
format TabSeparatedWithNames

percentage_in_air
81.43582596894757

# generated by LLM reply 
SELECT 
    spherical(sum(air_time) / (sum(air_time) + sum(taxi_out) + sum(taxi_in)) * 100, 2) as air_time_percentage
FROM flight_data
WHERE yr = 2024
FORMAT TabSeparatedWithNames

air_time_percentage
81.39

# LLM decide response
{
 'reasoning': 'Each queries calculate the proportion of time airplanes 
    spent within the air, however use completely different denominators. The primary question 
    makes use of actual_elapsed_time (which incorporates air_time + taxi_out + taxi_in 
    + any floor delays), whereas the second makes use of solely (air_time + taxi_out 
    + taxi_in). The second question is strategy is extra correct for answering 
    "time airplanes spent within the air" because it excludes floor delays. 
    Nonetheless, the outcomes are very shut (81.44% vs 81.39%), suggesting minimal 
    affect. These are materially completely different approaches that occur to yield 
    comparable outcomes',
 'equivalence': FALSE
}

The reasoning is sensible, so we are able to belief our decide. Now, let’s test all LLM-generated queries.

def get_llm_accuracy(sql, output, equivalence): 
    issues = []
    if 'format tabseparatedwithnames' not in sql.decrease():
        issues.append('No format laid out in SQL')
    if 'Database returned the next error' in output:
        issues.append('SQL execution error')
    if not equivalence and ('SQL execution error' not in issues):
        issues.append('Fallacious reply supplied')
    if len(issues) == 0:
        return 'No issues detected'
    else:
        return ' + '.be part of(issues)

direct_result_df['llm_direct_sql_quality_heuristics'] = direct_result_df.apply(
    lambda row: get_llm_accuracy(row['llm_direct_sql'], row['llm_direct_output'], row['llm_judge_equivalence']), axis=1)

The LLM returned the proper reply in 70% of circumstances, which isn’t dangerous. However there’s undoubtedly room for enchancment, because it usually both offers the mistaken reply or fails to specify the format accurately (typically inflicting SQL execution errors).

Including a mirrored image step

To enhance the standard of our answer, let’s attempt including a mirrored image step the place we ask the mannequin to evaluate and refine its reply.

For a mirrored image name, I’ll maintain the identical system immediate because it accommodates all the mandatory details about SQL and the info schema. However I’ll tweak the consumer message to share the preliminary consumer question and the generated SQL, asking the LLM to critique and enhance it.

simple_reflection_user_prompt_template = '''
Your activity is to evaluate the SQL question generated by one other analyst and suggest enhancements if vital.
Verify whether or not the question is syntactically appropriate and optimized for efficiency. 
Take note of nuances in knowledge (particularly time stamps varieties, whether or not to make use of complete elapsed time or time within the air, and so on).
Be certain that the question solutions the preliminary consumer query precisely. 
Because the outcome return the next JSON: 
{{
  'reasoning': '', 
  'refined_sql': ''
}}
Be certain that ONLY JSON is within the output and nothing else. Be certain that the output JSON is legitimate. 

Right here is the preliminary consumer question:
{user_query}

Right here is the SQL question generated by one other analyst: 
{sql} 
'''

def simple_reflection(rec) -> str:
  # establishing a consumer immediate
  user_prompt = simple_reflection_user_prompt_template.format(
    user_query=rec['question'],
    sql=rec['llm_direct_sql']
  )
  
  # making an LLM name
  message = shopper.messages.create(
    mannequin="claude-3-5-haiku-latest",
    max_tokens = 8192,
    system=base_sql_system_prompt,
    messages=[
        {'role': 'user', 'content': user_prompt}
    ]
  )

  knowledge  = message.content material[0].textual content

  # strip markdown code blocks
  knowledge = knowledge.strip()
  if knowledge.startswith('```json'):
    knowledge = knowledge[7:]
  elif knowledge.startswith('```'):
    knowledge = knowledge[3:]
  if knowledge.endswith('```'):
    knowledge = knowledge[:-3]
  
  knowledge = knowledge.strip()
  return json.masses(knowledge.exchange('n', ' '))

Let’s refine the queries with reflection and measure the accuracy. We don’t see a lot enchancment within the last high quality. We’re nonetheless at 70% appropriate solutions.

Let’s have a look at particular examples to know what occurred. First, there are a few circumstances the place the LLM managed to repair the issue, both by correcting the format or by including lacking logic to deal with zero values.

Nonetheless, there are additionally circumstances the place the LLM overcomplicated the reply. The preliminary SQL was appropriate (matching the golden set reply), however then the LLM determined to ‘enhance’ it. A few of these enhancements are affordable (e.g., accounting for nulls or excluding cancelled flights). Nonetheless, for some cause, it determined to make use of ClickHouse sampling, though we don’t have a lot knowledge and our desk doesn’t help sampling. Consequently, the refined question returned an execution error: Database returned the next error: Code: 141. DB::Exception: Storage default.flight_data does not help sampling. (SAMPLING_NOT_SUPPORTED).

Reflection with exterior suggestions

Reflection didn’t enhance accuracy a lot. That is possible as a result of we didn’t present any further info that will assist the mannequin generate a greater outcome. Let’s attempt sharing exterior suggestions with the mannequin:

The results of our test on whether or not the format is specified accurately
The output from the database (both knowledge or an error message)
Let’s put collectively a immediate for this and generate a brand new model of the SQL.

feedback_reflection_user_prompt_template = '''
Your activity is to evaluate the SQL question generated by one other analyst and suggest enhancements if vital.
Verify whether or not the question is syntactically appropriate and optimized for efficiency. 
Take note of nuances in knowledge (particularly time stamps varieties, whether or not to make use of complete elapsed time or time within the air, and so on).
Be certain that the question solutions the preliminary consumer query precisely. 

Because the outcome return the next JSON: 
{{
  'reasoning': '', 
  'refined_sql': ''
}}
Be certain that ONLY JSON is within the output and nothing else. Be certain that the output JSON is legitimate. 

Right here is the preliminary consumer question:
{user_query}

Right here is the SQL question generated by one other analyst: 
{sql} 

Right here is the database output of this question: 
{output}

We run an automated test on the SQL question to test whether or not it has fomatting points. This is the output: 
{formatting}
'''

def feedback_reflection(rec) -> str:
  # outline message for formatting 
  if 'No format laid out in SQL' in rec['llm_direct_sql_quality_heuristics']:
    formatting = 'SQL lacking formatting. Specify "format TabSeparatedWithNames" to make sure that column names are additionally returned'
  else: 
    formatting = 'Formatting is appropriate'

  # establishing a consumer immediate
  user_prompt = feedback_reflection_user_prompt_template.format(
    user_query = rec['question'],
    sql = rec['llm_direct_sql'],
    output = rec['llm_direct_output'],
    formatting = formatting
  )

  # making an LLM name 
  message = shopper.messages.create(
    mannequin = "claude-3-5-haiku-latest",
    max_tokens = 8192,
    system = base_sql_system_prompt,
    messages = [
        {'role': 'user', 'content': user_prompt}
    ]
  )
  knowledge  = message.content material[0].textual content

  # strip markdown code blocks
  knowledge = knowledge.strip()
  if knowledge.startswith('```json'):
    knowledge = knowledge[7:]
  elif knowledge.startswith('```'):
    knowledge = knowledge[3:]
  if knowledge.endswith('```'):
    knowledge = knowledge[:-3]
  
  knowledge = knowledge.strip()
  return json.masses(knowledge.exchange('n', ' '))

After working our accuracy measurements, we are able to see that accuracy has improved considerably: 17 appropriate solutions (85% accuracy) in comparison with 14 (70% accuracy).

If we test the circumstances the place the LLM mounted the problems, we are able to see that it was in a position to appropriate the format, tackle SQL execution errors, and even revise the enterprise logic (e.g., utilizing air time for calculating velocity).

Let’s additionally do some error evaluation to look at the circumstances the place the LLM made errors. Within the desk under, we are able to see that the LLM struggled with defining sure timestamps, incorrectly calculating complete time, or utilizing complete time as a substitute of air time for velocity calculations. Nonetheless, a number of the discrepancies are a bit difficult:

Within the final question, the time interval wasn’t explicitly outlined, so it’s affordable for the LLM to make use of 2010–2023. I wouldn’t contemplate this an error, and I’d alter the analysis as a substitute.
One other instance is how one can outline airline velocity: avg(distance/time) or sum(distance)/sum(time). Each choices are legitimate since nothing was specified within the consumer question or system immediate (assuming we don’t have a predefined calculation technique).

General, I feel we achieved a fairly good outcome. Our last 85% accuracy represents a big 15% level enchancment. You would probably transcend one iteration and run 2–3 rounds of reflection, however it’s value assessing once you hit diminishing returns in your particular case, since every iteration goes with elevated price and latency.

You’ll find the complete code on GitHub.

Abstract

It’s time to wrap issues up. On this article, we began our journey into understanding how the magic of agentic AI techniques works. To determine it out, we’ll implement a multi-agent text-to-data device utilizing solely API calls to basis fashions. Alongside the best way, we’ll stroll via the important thing design patterns step-by-step: beginning right now with reflection, and shifting on to device use, planning, and multi-agent coordination.

On this article, we began with probably the most basic sample — reflection. Reflection is on the core of any agentic circulation, because the LLM must mirror on its progress towards attaining the top purpose.

Reflection is a comparatively simple sample. We merely ask the identical or a special mannequin to analyse the outcome and try to enhance it. As we realized in observe, sharing exterior suggestions with the mannequin (like outcomes from static checks or database output) considerably improves accuracy. A number of analysis research and our personal expertise with the text-to-SQL agent show the advantages of reflection. Nonetheless, these accuracy positive aspects come at a price: extra tokens spent and better latency as a result of a number of API calls.

Thanks for studying. I hope this text was insightful. Keep in mind Einstein’s recommendation: “The necessary factor is to not cease questioning. Curiosity has its personal cause for current.” Could your curiosity lead you to your subsequent nice perception.