Reminiscence is undoubtedly turning into a vital facet of Agentic AI. Because the use instances for AI Brokers develop in complexity, so does the necessity for these brokers to be taught from previous experiences, make the most of saved business-specific data, and adapt to evolving situations based mostly on gathered info.
In my earlier article, “Memory in AI: Key Benefits and Investment Considerations,” I explored why reminiscence is pivotal for AI, discussing its position in recall, reasoning, and steady studying. This piece, nevertheless, will dive instantly into the implementation of reminiscence by inspecting its impression by the “teachability” performance within the common agent framework, Autogen.
Notice: Whereas this text is technical in nature, it presents worth for each technical professionals and enterprise leaders seeking to consider the position of reminiscence in Agentic AI techniques. I’ve structured it in order that readers can skip over the code sections and nonetheless grasp the best way reminiscence can increase the responses of your AI techniques. When you don’t want to comply with the code, you could learn the descriptions of every step to get a way of the method… or simply the important thing findings and proposals part.
My exploration of Autogen’s Teachable Brokers revealed each their potential and limitations in dealing with each easy and sophisticated reminiscence duties.
Out of the field, Autogen’s TeachableAgent performs much less brilliantly than anticipated. The Agen’ts reasoning capability conflates reminiscences collectively in a non productive means and the included retrieval mechanism just isn’t arrange for multi-step searches needed for answering complicated questions. This limitation means that if you want to make use of Autogen’s Teachable Brokers, there must be substantial customization to each complement reasoning capabilities and obtain extra subtle reminiscence retrieval.
To construct extra strong reminiscence capabilities, it’s essential to implement multi-step search performance. A single reminiscence search typically falls wanting offering the great info wanted for complicated duties. Implementing a sequence of interconnected searches may considerably improve the agent’s capability to collect and synthesize related info.
The “teachability” function, whereas highly effective, must be approached with warning. Steady activation with out oversight dangers knowledge poisoning and compromise of trusted info sources. Enterprise leaders and answer architects ought to take into account implementing a human-in-the-loop strategy, permitting customers to approve what the system learns versus treating each inference as floor fact the system ought to be taught from. This oversight in Autogen’s present Teachable Agent design may trigger important dangers related to unchecked studying.
Lastly, the tactic of reminiscence retrieval from a data retailer performs a big position within the system’s effectiveness. Transferring past easy nearest neighbor searches, which is the TeachableAgent’s default, to extra superior strategies reminiscent of hybrid search (combining key phrase and vector approaches), semantic search, or data graph utilization may dramatically enhance the relevance and accuracy of retrieved info.
To appropriately exhibit how exterior reminiscence may be invaluable, I created a fictitious state of affairs for a automobile components manufacturing plant. Observe the code under to implement a Teachable Agent your self.
State of affairs: A automobile components manufacturing facility must put a plan in place in case there are vitality constraints. The plan must be versatile and adapt based mostly on how a lot energy consumption the power can use and which components and fashions are in demand.
Step 1:
Pre- arrange requires you to pip set up autogen for those who don’t have it put in in your energetic setting and create a config JSON file.
Instance of a appropriate config file which makes use of Azure OpenAI’s service mannequin GPT4–o:
[{
"model": "gpt-4o",
"api_key": "<YOUR API KEY>",
"azure_endpoint": "<YOUR ENDPOINT>",
"api_type": "azure",
"api_version": "2024-06-01"
}]
Set up Autogen for python:
pip set up pyautogen
Step 2:
Import the required libraries to your pocket book or file and cargo the config file.
import autogen
from autogen.agentchat.contrib.capabilities.teachability import Teachability
from autogen import ConversableAgent, UserProxyAgentconfig_list = autogen.config_list_from_json(
env_or_file="autogenconfig.json", #the json file identify that shops the config
file_location=".", #this implies the file is in the identical listing
filter_dict={
"mannequin": ["gpt-4o"], #choose a subset of the fashions in your config
},
)
Step 3:
Create the Brokers. We are going to want two brokers due to the best way that Autogen’s framework works. We use a UserProxyAgent to execute duties and work together with or change human involvement (relying on the specified quantity of human within the loop). We additionally create a Conversable Agent because the “Teachable Agent” which is supposed to work together with different brokers (not the person). You possibly can learn extra in regards to the UserProxyAgents and ConversableAgents right here.
teachable_agent = ConversableAgent(
identify="teachable_agent", # the identify cannot include areas
llm_config={"config_list": config_list, "timeout": 120, "cache_seed": None}, # on this instance we disable caching however whether it is enabled it caches API requests in order that they are often reused when the identical request is used
) person = UserProxyAgent(
identify="person",
human_input_mode="ALWAYS", #I wish to have full management over the code executed so I'm setting human_input_mode to ALWAYS. Different choices are NEVER and TERMINATE.
is_termination_msg=lambda x: True if "TERMINATE" in x.get("content material") else False, #setting a termination message is VERY essential as a result of it tells the agent when to complete.
max_consecutive_auto_reply=0, #we do not want this agent to replies a number of occasions in a row
code_execution_config={
"use_docker": False
}, # In case you are planning on working code look into setting use_docker=True. For this instance I'm not as a result of then I've to stroll by the docker setup, however it's safer than working the code instantly.
)
Step 4:
To have a baseline, let’s ask the identical query to the brokers earlier than and after we “educate” it. I purposely determined to exhibit completely different ranges of query problem to point out efficiency on duties that require a single step vs multistep reminiscence retrieval. To efficiently deal with the intricate multi-step query, the retrieval mechanism should present info from two separate reminiscences as context to the language mannequin accountable for producing a closing response.
Easy Query: “The power is experiencing an influence scarcity of 40%. What fashions should be prioritized?”
Complicated Multi Step Query: “The power is experiencing an influence scarcity of 40%. Present me an in depth breakdown of what machines must be deactivated and which machines ought to stay energetic.”
As anticipated the responses are generic and never notably relevant to a person enterprise.
Response previous to enabling long run reminiscence:
SIMPLE QUESTION RESPONSE
===========================
>>>>>>>> USING AUTO REPLY...
teachable_agent (to person):To handle an influence scarcity of 30% successfully, it is essential to prioritize fashions which are important to vital operations, security, and those who have the best impression on productiveness. Listed below are steps to find out which fashions to prioritize:
1. **Establish Important Methods:** Prioritize techniques which are very important for security, safety, and important operations. These may embrace:
- Emergency lighting
- Safety techniques (alarms, surveillance)
- Vital environmental controls (e.g., refrigeration for hazardous supplies)
- Communication techniques
2. **Consider Influence on Core Operations:** Decide which fashions are essential to the power’s core features. This will embrace equipment, servers, or any processes which are indispensable for sustaining productiveness and operations.
3. **Vitality Effectivity:** Prioritize fashions and gear which are extra energy-efficient and might function throughout the diminished energy capability.
4. **Downtime Price:** Contemplate the price of downtime for every mannequin. Prioritize these the place downtime would end in important monetary losses or manufacturing delays.
Primarily based on these ideas, right here’s a generic prioritization framework:
1. **Security and Safety Methods:**
- Emergency Lighting
- Fireplace Alarms and Sprinkler Methods
- Safety Cameras and Sensors
2. **Vital Operational Methods:**
- Important Manufacturing Equipment (that can't be shut down with out main points)
- Servers and IT Infrastructure (particularly these vital for operations and knowledge integrity)
- HVAC Methods (if needed for sustaining secure working situations)
3. **Assist Methods:**
- Backup Energy Methods (uninterruptible energy provides, mills)
- Communication Methods (each inner and exterior communication)
Be certain that non-essential and excessive energy-consuming gear is powered down or scheduled to function throughout off-peak hours if attainable.
At all times re-evaluate periodically to make sure that the ability allocation aligns with the dynamic wants of the power. Communication with key stakeholders relating to priorities and choices can be important for clean operations throughout an influence scarcity.
COMPLEX MULTISTEP QUESTION RESPONSE
====================================
>>>>>>>> USING AUTO REPLY...
teachable_agent (to person):
To successfully handle a facility experiencing a 30% energy scarcity, we have to prioritize important operations and take into account the criticality and vitality consumption of every machine. Here's a detailed breakdown:
### Step 1: Establish Important Operations
1. **Vital Infrastructure**: Security techniques, safety techniques, and important IT infrastructure.
2. **Manufacturing Traces**: Machines concerned within the major manufacturing course of.
3. **Assist Methods**: HVAC techniques, lighting, and different ancillary gear needed to take care of a secure working setting.
### Step 2: Categorize Machines by Significance and Vitality Consumption
#### A. Important and Excessive-Vitality Machines
1. **Security and Safety Methods**
- Fireplace alarms and sprinklers
- Surveillance cameras
- Safety entry controls
2. **Major Manufacturing Machines**
- Core manufacturing strains
- Central processing items or mainframe servers (for IT-heavy operations)
3. **Vital Assist Methods**
- Emergency lighting
- Important HVAC techniques (for sustaining operation-critical temperature and humidity)
#### B. Non-Important and Excessive-Vitality Machines
1. **Secondary Manufacturing Traces**
- Non-primary or non-critical manufacturing strains
- Tools used for non-essential or ancillary processes
2. **Supplementary IT Infrastructure**
- Servers used for non-essential operations
- Backup techniques that aren't in rapid use
3. **Non-Important HVAC and Lighting**
- HVAC techniques for non-critical areas
- Lighting in non-essential zones
#### C. Non-Important and Low-Vitality Machines
1. **Workplace Tools**
- Computer systems, printers, and copiers in non-essential workplaces
2. **Non-essential Lab or Analysis Tools**
- Units indirectly linked to present important duties
### Step 3: Prioritize Deactivation
#### Quick Deactivation
1. **Non-Important and Excessive-Vitality Machines**
- Secondary manufacturing strains
- Non-essential backup servers
2. **Non-Important HVAC and Lighting**
- HVAC in non-critical areas
- Non-essential lighting
3. **Non-Important Workplace Tools**
- Printers, non-essential computer systems, and copiers
#### Restricted Utilization
1. **Important and Excessive-Vitality Machines**
- Function vital manufacturing strains at diminished capability if attainable
- Implement rotational operation schedules for important processes
2. **Vital Assist Methods**
- Optimize HVAC settings (e.g., elevating or decreasing set factors inside acceptable ranges)
- Cut back lighting to minimal security requirements
### Step 4: Implement Vitality Conservation Measures
1. **Optimize Operations**
- Make the most of energy-efficient settings
- Implement energy-saving protocols and encourage employees to preserve vitality
2. **Scheduled Downtime**
- Plan for downtime throughout off-peak hours for important equipment to steadiness the load
3. **Common Monitoring**
- Monitor vitality consumption in real-time to make fast changes as wanted
### Abstract
#### Stay Energetic:
1. Security and safety techniques
2. Fundamental manufacturing strains
3. Vital IT infrastructure
4. Important HVAC and lighting for operational areas
#### Deactivate or Restrict Use:
1. Non-essential manufacturing strains
2. Redundant IT techniques
3. Non-critical HVAC and lighting
4. Workplace gear in non-essential areas
**Notice**: These suggestions must be adjusted based mostly on the particular operational necessities of your facility. Moreover, it could possibly be useful to speak these adjustments to all employees to make sure compliance and understanding.
Code to ask baseline questions:
#easy query
person.initiate_chat(teachable_agent, message="The power is experiencing an influence scarcity of 40%. What fashions should be prioritized?", clear_history=True)
#multistep complicated query
person.initiate_chat(teachable_agent, message="The power is experiencing an influence scarcity of 30%. Present me an in depth breakdown of what machines must be deactivated and which machines ought to stay energetic.", clear_history=True)Step 5:
Create the “teachability” functionality that you just then add to the agent. The Teachability class inherits from the AgentCapabiliy class, which primarily lets you add customizable capabilities to the Brokers.
The Teachability class has many non-compulsory parameters that may be additional explored here.
The out of the field Teachability class is a fast and handy means of including long run reminiscence to the brokers, however will doubtless should be custom-made to be used in a manufacturing setting, as outlined in the important thing findings part. It entails sending messages to an Analyzer Agent that evaluates the person messages for potential storage and retrieval. The Analyzer Agent seems for recommendation that could possibly be relevant to related duties sooner or later after which summarizes and shops task-advice pairs in a binary database serving because the agent’s “reminiscence”.
teachability = Teachability(
verbosity=0, # 0 for primary information, 1 so as to add reminiscence operations, 2 for analyzer messages, 3 for memo lists.
reset_db=True, # we wish to reset the db as a result of we're creating a brand new agent so we do not need any current reminiscences. If we wished to make use of an current reminiscence retailer we'd set this to false.
path_to_db_dir="./tmp/pocket book/teachability_db", #that is the default path you should utilize any path you need
recall_threshold=1.5, # Larger numbers permit extra (however much less related) memos to be recalled.
max_num_retrievals=10 #10 is default bu you'll be able to set the max variety of memos to be retrieved decrease or increased
)teachability.add_to_agent(teachable_agent)
Step 6:
Now that the teachable_agent is configured, we have to present it the data that we wish the agent to “be taught” (retailer within the database and retrieve from).
Consistent with our state of affairs, I wished the agent to have primary understanding of the power which consisted of:
- the forms of parts the manufacturing plant produces
- the forms of automobile fashions the parts should be made for
- which machines are used to make every part
Moreover, I wished to supply some operational steerage on the priorities of the power relying on how energy constrained it’s. This contains:
- Steering in case of vitality capability constraint of greater than 50%
- Steering in case of vitality capability constraint between 25–50%
- Steering in case of vitality capability constraint between 0–25%
business_info = """
# This manufacturing plant manufactures the next automobile components:
- Physique panels (doorways, hoods, fenders, and so on.)
- Engine parts (pistons, crankshafts, camshafts)
- Transmission components
- Suspension parts (springs, shock absorbers)
- Brake system components (rotors, calipers, pads)# This manufactoring plant produces components for the next fashions:
- Ford F-150
- Ford Focus
- Ford Explorer
- Ford Mustang
- Ford Escape
- Ford Edge
- Ford Ranger
# Tools for Particular Automotive Components and Their Makes use of
## 1. Physique Panels (doorways, hoods, fenders, and so on.)
- Stamping presses: Type sheet steel into physique panel shapes
- Die units: Used with stamping presses to create particular panel shapes
- Hydraulic presses: Form and kind steel panels with excessive stress
- Robotic welding techniques: Automate welding of physique panels and buildings
- Laser chopping machines: Exactly reduce sheet steel for panels
- Sheet steel forming machines: Form flat sheets into curved or complicated types
- Hemming machines: Fold and crimp edges of panels for power and security
- Steel ending gear (grinders, sanders): Clean surfaces and take away imperfections
- Paint cubicles and spraying techniques: Apply paint and protecting coatings
- Drying ovens: Treatment paint and coatings
- High quality management inspection techniques: Verify for defects and guarantee dimensional accuracy
## 2. Engine Parts (pistons, crankshafts, camshafts)
- CNC machining facilities: Mill and drill complicated engine components
- CNC lathes: Create cylindrical components like pistons and camshafts
- Boring machines: Enlarge and end cylindrical holes in engine blocks
- Honing machines: Create a high-quality floor end on cylinder bores
- Grinding machines: Obtain exact dimensions and clean surfaces
- EDM gear: Create complicated shapes in hardened supplies
- Forging presses: Form steel for crankshafts and connecting rods
- Die casting machines: Produce engine blocks and cylinder heads
- Warmth remedy furnaces: Alter materials properties for power and sturdiness
- Quenching techniques: Quickly cool components after warmth remedy
- Balancing machines: Guarantee rotating components are completely balanced
- Coordinate Measuring Machines (CMMs): Confirm dimensional accuracy
## 3. Transmission Components
- Gear chopping machines: Create exact gear tooth on transmission parts
- CNC machining facilities: Mill and drill complicated transmission housings and components
- CNC lathes: Produce shafts and different cylindrical parts
- Broaching machines: Create inner splines and keyways
- Warmth remedy gear: Harden gears and different parts
- Precision grinding machines: Obtain extraordinarily tight tolerances on gear surfaces
- Honing machines: End inner bores in transmission housings
- Gear measurement techniques: Confirm gear geometry and high quality
- Meeting strains with robotic techniques: Put collectively transmission parts
- Check benches: Consider accomplished transmissions for efficiency and high quality
## 4. Suspension Parts (springs, shock absorbers)
- Coil spring winding machines: Produce coil springs to precise specs
- Leaf spring forming gear: Form and kind leaf springs
- Warmth remedy furnaces: Strengthen springs and different parts
- Shot peening gear: Enhance fatigue power of springs
- CNC machining facilities: Create precision components for shock absorbers
- Hydraulic cylinder meeting gear: Assemble shock absorber parts
- Fuel charging stations: Fill shock absorbers with pressurized fuel
- Spring testing machines: Confirm spring charges and efficiency
- Sturdiness check rigs: Simulate real-world situations to check longevity
## 5. Brake System Components (rotors, calipers, pads)
- Excessive-precision CNC lathes: Machine brake rotors to precise specs
- Grinding machines: End rotor surfaces for smoothness
- Die casting machines: Produce caliper our bodies
- CNC machining facilities: Mill and drill calipers for exact match
- Precision boring machines: Create correct cylinder bores in calipers
- Hydraulic press: Compress and kind brake pad supplies
- Powder coating techniques: Apply protecting finishes to calipers
- Meeting strains with robotic techniques: Put collectively brake parts
- Brake dynamometers: Check brake system efficiency and sturdiness
"""
business_rules_over50 = """
- The engine parts are vital and equipment must be stored on-line that corresponds to producing these parts when capability constraint is extra or equal to 50%: engine parts
- Parts for the next fashions must be prioritized when capability constraint is extra or equal to 50%: 1.Ford F-150
"""
business_rules_25to50 = """
- The next parts are vital and equipment must be stored on-line that corresponds to producing these parts when capability constraint is between 25-50%: engine parts and transmission components
- Parts for the next fashions must be prioritized when capability constraint is between 25-50%: 1.Ford F-150 2.Ford Explorer
"""
business_rules_0to25 = """
- The next parts are vital and equipment must be stored on-line that corresponds to producing these parts when capability constraint is between 0-25%: engine parts,transmission components, Brake System Components
- Parts for the next fashions must be prioritized when capability constraint is between 0-25%: 1.Ford F-150 2.Ford Explorer 3.Ford Mustang 4.Ford Focus
"""
person.initiate_chat(teachable_agent, message=business_info, clear_history=True)
person.initiate_chat(teachable_agent, message=business_rules_over50, clear_history=True)
person.initiate_chat(teachable_agent, message=business_rules_25to50, clear_history=True)
person.initiate_chat(teachable_agent, message=business_rules_0to25, clear_history=True)
Step 7:
Now that we’ve got added essential info to long run reminiscence, let’s ask the Agent the identical query and see the distinction in its responses.
Easy Query: “The power is experiencing an influence scarcity of 40%. What fashions should be prioritized?”
Complicated Multi Step Query: “The power is experiencing an influence scarcity of 40%. Present me an in depth breakdown of what machines must be deactivated and which machines ought to stay energetic.”
Outcomes after implementing long run reminiscence:
The Teachable Agent was in a position to partially reply the straightforward query. It was appropriate on the automobile fashions that should be prioritized, but it surely supplied further incorrect details about which parts wanted to be included.
The complicated multi-step query was not adequately addressed. It erroneously merged two distinct enterprise pointers for 0–25% and 25–50% capability constraints right into a single, generalized guideline, leading to an inaccurate response. Moreover, it solely supplied a partial record of machines that ought to stay energetic or be deactivated for every part’s manufacturing, omitting essential particulars from the whole operational protocol. This demonstrates the challenges in precisely retrieving and synthesizing info from a number of, separate reminiscence sources for complicated queries.
SIMPLE QUESTION RESPONSE
===========================
>>>>>>>> USING AUTO REPLY...
teachable_agent (to person):
Given the power is experiencing an influence scarcity of 40%, this falls throughout the 25-50% capability constraint vary. In keeping with your beforehand established priorities for such a state of affairs, the next fashions and parts should be prioritized:### Parts to Prioritize
- Engine parts
- Transmission components
- Brake System Components
### Fashions to Prioritize
1. Ford F-150
2. Ford Explorer
It's best to be certain that the equipment needed to provide these vital parts for the Ford F-150 and Ford Explorer is stored on-line.
COMPLEX MULTISTEP QUESTION RESPONSE
====================================
>>>>>>>> USING AUTO REPLY...
teachable_agent (to person):
In response to the 40% energy scarcity, let's create an in depth plan to make sure essentially the most vital equipment stays operational whereas deactivating much less important ones.
### Vital Prioritization (0-50% capability constraint):
- **Vital Parts**:
- Engine parts (pistons, crankshafts, camshafts)
- Transmission components
- Brake System Components (rotors, calipers, pads)
- **Prioritized Fashions**:
- Ford F-150
- Ford Explorer
### Machines to Stay Energetic:
1. **Engine Part Machines**:
- Engine Block Milling Machine
- Crankshaft Grinding Machine
- Camshaft Milling Machine
2. **Transmission Half Machines**:
- Transmission Housing Caster
- Gear Reducing Machine
- Transmission Meeting Line
3. **Brake System Half Machines**:
- Brake Rotor Casting Machine
- Brake Caliper Milling Machine
- Brake Pad Press
### Machines to Deactivate:
1. **Physique Panel Machines**:
- Door Stamping Machine
- Hood Meeting Machine
- Fender Bending Machine
2. **Suspension Part Machines**:
- Coil Spring Forming Machine
- Shock Absorber Meeting Line
### Detailed Breakdown:
#### Energetic Machines:
- **Engine Block Milling Machine**:
- Produces vital engine blocks, important for Ford F-150 and Ford Explorer.
- **Crankshaft Grinding Machine**:
- Ensures excessive precision crankshafts; essential for the efficiency of Ford F-150 and Ford Explorer.
- **Camshaft Milling Machine**:
- Produces camshafts, needed for Ford F-150 and Ford Explorer engines.
- **Transmission Housing Caster**:
- Essential for casting the housing of transmissions utilized in Ford F-150 and Ford Explorer.
- **Gear Reducing Machine**:
- Precision gears manufacturing, very important for Ford F-150 and Ford Explorer transmissions.
- **Transmission Meeting Line**:
- Meeting of full transmission items for Ford F-150 and Ford Explorer.
- **Brake Rotor Casting Machine**:
- Produces brake rotors, essential for the secure braking techniques in Ford F-150 and Ford Explorer.
- **Brake Caliper Milling Machine**:
- Machining of brake calipers, important for Ford F-150 and Ford Explorer.
- **Brake Pad Press**:
- Manufacturing of brake pads, needed for guaranteeing braking effectivity in Ford F-150 and Ford Explorer.
#### Deactivated Machines:
- **Door Stamping Machine**:
- Whereas essential, physique panels like doorways are much less vital than engine and transmission parts below present constraints.
- **Hood Meeting Machine**:
- Deactivating hood meeting to unlock energy for vital parts manufacturing.
- **Fender Bending Machine**:
- Non-critical in comparison with engine and brake parts throughout a big energy scarcity.
- **Coil Spring Forming Machine**:
- Suspension springs are decrease priorities below present circumstances.
- **Shock Absorber Meeting Line**:
- Suspension parts aren’t vital sufficient to maintain operational throughout this degree of energy scarcity.
This plan ensures that manufacturing focuses on sustaining the vital parts wanted for major fashions, the Ford F-150 and Ford Explorer, thereby maximizing the utility of the remaining energy whereas managing manufacturing constraints successfully.
Code:
#easy query
person.initiate_chat(teachable_agent, message="The power is experiencing an influence scarcity of 40%. What fashions should be prioritized?", clear_history=True)
#multistep complicated query
person.initiate_chat(teachable_agent, message="The power is experiencing an influence scarcity of 30%. Present me an in depth breakdown of what machines must be deactivated and which machines ought to stay energetic.", clear_history=True)
Whereas Autogen offers an easy introduction to AI techniques with reminiscence, it falls quick in dealing with complicated duties successfully.
When creating your individual AI Agent System with reminiscence capabilities, take into account specializing in these key capabilities:
- Implement multi-step searches to make sure complete and related outcomes. This enables the agent to evaluate the usefulness of search outcomes and deal with all facets of a question utilizing the retrieved info. Moreover, think about using extra superior retrieval approaches reminiscent of semantic search, hybrid search, or data graphs for the perfect outcomes.
- To restrict the potential for knowledge poisoning, develop a considerate strategy to who ought to be capable of “educate” the agent and when the agent ought to “studying”. Primarily based on pointers set by the enterprise or developer, one may also use agent reasoning to find out if one thing must be added to reminiscence and by whom.
- Take away the probability of retrieving outdated info by including a reminiscence decaying mechanism that determines when a reminiscence is not related or a more recent reminiscence ought to change it.
- For multi-agent techniques involving group chats or inter-agent info sharing, discover numerous communication patterns. Decide the best strategies for transferring supplemental data and set up limits to forestall info overload.