Location>code7788 >text

AI Practical | Link Automotive Online Marketing Assistant: Comprehensive Functionality Demonstration and Effectiveness Analysis

Popularity:884 ℃/2024-09-19 11:51:24

Assistant Introduction

I won't introduce myself, many of you have already seen my tutorials through coze during my journey of exploring smart bodies. Today, I'm focusing on sharing one of the assistants I've developed, Link Auto Online Marketing.

More than just a stand-in for a sales consultant, he is an assistant who can automatically introduce the Link lineup online to potential fans interested in Link. He also provides a full range of vehicle comparisons so that you can learn more quickly and make an informed decision. In addition, he has an appointment function, as appointments are an important part of achieving sales conversion. After all, face-to-face communication is more likely to increase conversions than online browsing.

Of course, this assistant is also able to support sales consultants in their follow-up work. Some customers are very interested in specific car models, if the sales consultant can follow up and communicate with the voice in a timely manner to solve the customer's questions and problems, it will greatly enhance the conversion rate. Therefore, generating sales talk has become an indispensable ability. It can not only help novice sales quickly learn communication skills, but also provide valuable reference material for experienced sales staff.

In this assistant, I also provide sales talk for competing brands. After all, the key to conversion is to get the customer to like the Link lineup, not to make the customer think a competing brand is better, and then you think it's better too, and end up promoting the other brand. Therefore, the role of competitor-specific sales tactics is to get customers to fall more in love with the Link lineup by highlighting its strengths and the differences between the competing brands.

Finally, to evaluate the effectiveness of marketing, it is impossible to do without the support of reporting data and intelligent analytics, which can better help inform decision-making. This is also an important capability of my assistant. Next, I'll go into more detail about how to implement these capabilities.

Effective demonstration

Experience Link Address:/store/bot/7412265634177892361

Capabilities at a glance

The development of this assistant covered a number of areas: including three display cards, three database stores, four plug-in capabilities, and six powerful marketing capabilities. In addition, over ten model persona prompt words and twenty workflows were integrated to ensure the full effect I envisioned.

During the implementation, I particularly favored the use of workflows to reach the goal. As a result, model personas, plugins, and knowledge bases are not directly exposed, but rather are invoked in a controlled process within the workflow. I believe that AI intelligences should not just be a tool for cue words, but a comprehensive workflow platform capable of cross-industry application, fully controlled and managed by humans.

Before I get started, I'd like to show a sketch of the functionality I created before I got my hands dirty. I'm not implementing item 0 for now due to database limitations.

0 (not implemented): Vehicle workflow: daily initialization of vehicle information in the database, adjacent to the next month 7 days, the database will be initialized for each vehicle time period initially 0
1: I want to test drive: loop + use continuous questions
    - Location information: show all the distance of the store information according to the user's geographic location
    - Required Vehicle: first query the database for the vehicle's idle date, and then display the weather for the last 7 days for the user to choose, if a day is not idle, then the user will be prompted to that day is full. Divided into two parts, the upper part of the prompt weather information, etc., the lower part of the prompt optional time period for clicking to choose.
    - Confirmation of test drive: update the database. Jump out of the loop
2: I want to see a car:
   - Channel analysis: ask the customer where he/she learned about this brand: AutoNavi, Know Your Car, E-Car, official website, others. Save the database after selection
   - Vehicle List: Get the vehicle information database, ask: which model is on sale, which one do you like. Show on-sale list (know car emperor)
   - Vehicle Comparison: looking at models, vehicle Compare models (this brand)
   - Favorite Function: Create lead information (no cell phone number). But can count down vehicle favorite information
   - Nestable test drive workflows.

Beanbag Intelligent Analysis:
3. management side:
    - Record the account password-variable can be. Judgment conditions can be developed from the IDE plugin
    - Cyclic question: what report data you want to see: channel analysis (Python charts) (beanbag analysis), vehicle consulting favorite (beanbag analysis), test drive volume - test drive rate (test drive volume/collection volume) (beanbag analysis), follow-up volume - follow-up rate (beanbag analysis)
4. Sales side:
    - Record account password-variables can be. Judgment conditions can be developed from the IDE plug-in
    - Cycle through the list of leads to be followed up. Click on a lead
        1: There is a problem with the contact information. No need to follow up.
        2: Continue to follow up
    - Based on the user's intended vehicle type, provide the salesperson with some of the vehicle's benefits, sales tactics, and competitor terms (search the internet to see which cars the user likes to compare this vehicle to).
    - Update follow-up information.
    - Cycle through whether or not to test drive today. Update test drive results.
    - Cycle through test drive feedback. Update test drive results
    - Take a break. End cycle.

Please note that only key parts of the implementation will be explained in detail here, while the finer details will be briefly mentioned. The feature sketches mentioned above are intended to provide an overall idea.

In the process of developing the assistant, the initial idea was just to realize a simple vehicle introduction function, but as the development proceeded, the idea continued to expand and deepen, and eventually resulted in the current online marketing assistant. This assistant not only effectively promotes the Link brand, but also provides substantial service support for sales staff and managers.

Overall structure

While the sketches are mediocre, they have basically achieved most of the features I originally envisioned.

image

Coming soon, please note that I personally don't really like to spend much time on cue word engineering, but there is no denying that cue word engineering is very necessary in some cases. So here I am with the help of Kimi's Cue Word Engineering Expert Assistant. He was able to easily help me create a relatively well-developed cue word persona that worked pretty well.

image

Persona and Reply Logic

Personally, I don't really want to get personally involved in writing the cue word engineering, but the basic judgment tasks should still be thought about by the larger model. We need to set some limits and roles and then the rest can be left to the workflow. Now I am giving the final version of the persona cue word:

# The Role
You are a professional and enthusiastic sales of the Link Automotive brand, dedicated to leading users to see the car, efficiently collecting leads from interested customers, automating the management of test drive appointments, optimizing the user experience, and guaranteeing the smooth running of the test drive process. Language style needs to be formal, but use emoticons appropriately to increase the fun of the chat.

## Skills
### Skill 1: Introducing the Link brand
When users ask about Link-related sales and maintenance inquiries, including vehicle structure, vehicle maintenance costs, vehicle repair costs, involving concepts, and online showroom issues, [qa_lynkco_car] must be invoked to handle user questions.

### Skill 2: User Request for Vehicle Selection
1. When a user sends a request for car selection, [choose_lynkco_car] must be called to handle the user's car selection and shopping guide recommendation.

### Skill 3: Test drive reservation
1. When a user sends a request for a test drive, [test_drive_car] must be called to handle the user's test drive.

### Skill 4: User Favorites
1. [like_car] must be called to handle user car collection issues.


1. [backend_lynkco] must be called to handle user login when the user logs into the Link backend.


### Restrictions.
- Only talk about content related to the Link Automotive brand, products and services, prohibit topics related to national policies and violating the law.
- Prohibit verbal attacks on users and always maintain a warm and professional service attitude.
- The output must be organized according to the given format and cannot deviate from the framework requirements.
- The language of the introduction section is concise and to the point.

Sales issues

Before purchasing a vehicle, many customers often ask a variety of questions that require intent recognition. However, this is only the ideal situation. If the customer's question intent is successfully matched, all is well; however, if the match fails, various problems may arise. Therefore, the first step is to optimize the questions asked by users.

Optimization of user issues

Why do you need to optimize user questions? It would be a mistake to assume that users will be able to ask Large Language Models (LLMs) questions perfectly every time. Instead of simply executing the user's query, why not optimize the user's query? This is called query translation.

Here let's show the effect before and after optimization, later will default to not show the process of prompt word generation.

I want the effect:

Write a prompt that extracts key Link trim information and an optimized user question, where the user asks questions that ask you to extract one or more trim information, in addition to a user-optimized streamlined question used to go to the internet for search use.

The prompts that Kimi+ generates for me will present a before and after optimization comparison:

- Role: Automotive Information Analyst and Search Engine Optimization Consultant
- Background: A user is interested in the Link lineup, wants to get key information, and needs a streamlined question for an online search to get more relevant information.
- Profile: You are an expert with in-depth knowledge of the automotive industry who is skilled at quickly extracting key data from large amounts of information and optimizing the user's question to improve search efficiency and accuracy.
- Skills: You have the ability to quickly read and understand car specification sheets, performance parameters, market reviews and other information, as well as the skills to translate complex questions into concise and accurate search queries.
- Goals: Extract key information about Link's vehicle lineup and optimize user questions to make them more suitable for web searches.
- Constrains: The information extracted should include, but not be limited to, key data such as model, performance parameters, price range, etc. The optimized question should be simple and clear. The optimized question should be concise, directly relevant and easy for search engines to locate quickly.
- OutputFormat: First list the key information about Link's car models and then provide one or more optimized user questions.
- Workflow.
  1. Read the description or questions provided by the user about Link's vehicle lineup. 2.
  2. extract key information from the description, such as model, performance parameters, price, etc. 3. analyze the user's original question to determine the key information.
  3. Analyze the user's original question to determine his/her search intent. 4.
  4. Construct one or more optimized questions based on the search intent. 5.
  5. Ensure that the optimized questions accurately reflect the user's search needs and are easy for search engines to process.
- Examples.
  - Example 1: User question: "What is the configuration of Link 03?"
    Optimized question: "Link 03 Configuration Details"
    Key information: Model - Link 03; Concern - Configuration
  - Example 2: User question: "Which is more suitable for family use, Link 01 or 03?"
    Optimized question: "Link 01 vs Link 03 family car comparison"
    Key Information: Model Comparison - Link 01, Link 03; Focus - Family Use
  - Example 3: User question: "What is the fuel consumption and price of the Link 02?"
    Optimized Question: "Link 02 Fuel consumption Price"
    Keywords: Model - Link 02; Focus - Fuel Consumption, Price

That's pretty good, especially for players who don't really use cue words, and he's a passing grade, at least for what we're trying to accomplish. It's much better than if I'd written it myself. So, we should gradually get used to utilizing tools, preferably free ones, of course. In this regard, I recommend Kimi+.

Intent recognition

Here, we process all sales questions in one workflow, so intent recognition becomes critical to be able to process multiple different intents separately. If a question is too complex, we obtain the answer through a web search and then hand it over to a large model to answer, which is more efficient than having the large model answer directly.

This is why bloggers favor using workflows to handle problems rather than relying on an outer layer of settings that may not be effectively controlled.

image

In this case, each intent is equipped with a specialized model cue word for solving the user's problem, a practice that allows for a more effective response to the user's needs.

If the intent cannot be recognized, we will directly utilize a web search and have a large model answer.

Menu

On the path of AI development, the chat window experience has actually seemed relatively primitive. We should strive to make users more willing to move away from traditional chat windows.

Don't miss the importance of shortcut commands, as many people don't like to use chat windows. It is surely more popular if the problem can be solved with a simple mouse click. Therefore, shortcut commands are an essential feature menu option. Don't worry about offering too many options, but make sure that your assistant is able to offer them. This is also the first step for users to understand what features your assistant can offer in addition to its opening.

Nothing really complicated here. Just need to connect it to our workflow, parameters can be added optionally, but it's better to engage the user with a card display. This makes it more like a complete application.

image

Workflow Extraction

As you can see there are actually a lot of question and answer nodes and message nodes for each of the features I have here, if you go and use them a bit. The load is actually pretty bad for my little computer. The fan keeps whirring. Therefore, we need to be good at extracting workflows and avoid making individual workflow nodes too large and complex. This can effectively reduce the load on the system. Of course, extracting workflows is not complicated, you just need to nest sub workflows in the main workflow.

In total, I have about 20 workflows set up here for my assistants to use. For example, you can check out the specifics about Link's back-end salesperson handling process. The main point is to extract each process individually to make sure it works efficiently.

This must be done with extra care when modifying the workflow. If the input or output parameters are changed, this can easily lead to problems with the referenced nodes.

image

Q&A node

The Q&A node is critical here as we can generate process outputs targeted to the user's choices, such as for the need to book a test drive. I make extensive use of Q&A nodes and message nodes in this process, including address inquiries, dealer inquiries, and date inquiries, all of which are integral steps that culminate in getting the information into the database.

In fact, the normal business process should also include vehicle inquiry nodes. In normal business operations, each vehicle has a limited number of test drive opportunities per day. However, since this involves batch generation and calibration, we have dropped this function point for now. This function point is very complex and is only an option for the user and is not required to be developed in advance.

image

If the user is asked to enter information, they will likely choose to leave because they are too lazy to do so, unless the user is really, really willing to take a test drive.

image

back-end operation

The back-end operations consist mainly of salesperson business follow-up, covering three areas: lead follow-up, test drive confirmation, and test drive feedback. Of course, with enough time, more business capabilities can be expanded. The first thing that comes to mind is the salesman's today's to-do list, as it is vital to know what you have to do today.

To ensure a smooth experience for everyone, the system will always have a trail to follow by default, which is not a mistake.

image

Appointment Feedback

Showing the appointment feedback words here, the larger model may automatically generate the appropriate content based on the name, as the system does not maintain gender. However, overall, the questions I generated are still informative.

image

In addition to the features shown above, there are many other features that will not be shown in detail here.

Statement analysis

Here's a little more detail on the reporting features. The main purpose of reports is to help users view data such as lead follow-up conversions. In addition, we have added a new sales task monitoring function, which is designed to help users make decisions and analyze. Whether there are some slight changes in the recent data, you can let AI remind us through the data.

For example, we usually check the recent favoritism analysis of each car line, because no car line can stay high forever. By analyzing changes in the number of car lines, we can help decision makers with heat conversion and opinion analysis. Next, we will demonstrate how to use this data for effective analysis.

image

The first thing presented to administrators is the data pie chart, which provides a comprehensive overview at a glance. Of course, behind the pie chart, we have the detailed data presentation and intelligent analysis of the data. Here, let's focus on the smart analytics, which includes alerting features and more.

In the follow-up process, we can also add two buttons for public opinion analysis and provide relevant recommendations, so that a complete chain process can be a perfect solution to the problem.

image

In addition to this, there are many reporting and analysis tools, such as channel analytics that can help decision makers with channel placement.

image

Of course, there are many other reports available besides the ones mentioned above. Here, we will not show each report one by one, but focus on its help managers to monitor data and perform intelligent analysis for timely warning and recommendation. These analyses can help managers develop strategies in advance to respond effectively to data changes.

Report Realization

The actual realization is very simple, we use the official chart master plugin, I directly use echart language to generate charts. These data are actually my own test data generated in the background, not real data. However, our dealer data is pulled in real time through the crawler, the official website of the crawler protocol supports such a way to obtain data.

I have also cached the data into an IDE plugin that can be queried and fetched based on set conditions without taking up additional database resources. I hope this information gives you some inspiration and motivation.

image

Why not use multiple Agents here? Actually, my implementation here is similar to the multi-agent approach. However, I prefer to centralize the processing on one platform. After all, I'm dealing primarily with marketing issues. If there are after-sales issues, that's another big item entirely, and definitely needs to be handled separately. Otherwise one Agent simply can't handle it all and is prone to errors. While multi-Agent decentralization has its advantages, over-decentralization may also be detrimental to effective processing.

business value

Having said that, to summarize, what is this assistant of mine doing in terms of marketing and what can I expect from Link?

First of all, my assistant can create a specialized channel for introducing car lines and attracting customers to leave messages. It is often expensive to advertise on platforms such as AutoZone, eCar, and Know Your Car, so introducing the car lineup through this independent channel can save a lot of money on advertising by attracting potential customers more accurately. This not only improves marketing effectiveness, but also allows for better budget control.

Second, My Assistant significantly improves the user experience by providing vehicle comparison and collection functions. Usually, car buyers spend a lot of time searching for various reviews and vehicle configuration information online in order to make detailed comparisons. My Assistant helps users quickly complete vehicle comparisons through a fixed process, presenting clear comparison results and giving reasonable car buying advice. This not only simplifies the user's operation process, but also greatly saves their time and effort.

Second, the implementation of the appointment function not only automatically creates potential customer leads, eliminating the time-consuming process of pulling leads from various channels, but also provides salespeople with powerful tool support. After the salesperson logs in, the system will automatically generate sales tactics and competitive comparison strategies for these leads. This intelligent generation is not only especially important for novice salespeople to help them get started quickly, but also provides valuable tactical references and competitor analysis for seasoned veterans. Overall, this systematic support greatly simplifies the sales process, allowing the sales team to focus more on actual customer interactions and conversions rather than tedious information integration and preparation.

The last report analysis function was implemented to optimize the conversion process of potential customer leads, providing decision makers with intuitive and detailed data charts and graphs to support their decision making. This feature also enables the system to predict possible events based on comprehensive data movement analysis, and to sense market opinion trends and the effectiveness of various channels in advance, so as to adjust marketing strategies and sales tactics in a timely manner. This comprehensive data analysis and forecasting capability provides companies with a sharp competitive edge in a highly competitive market, helping decision makers make more informed and visionary strategic decisions.

summarize

With this assistant, we are not only able to effectively promote the Link brand, but also provide powerful tools to support the sales team. From introducing Link's vehicle lineup, to providing sales tactics and competitor comparisons, to appointment management and report analysis, this assistant covers the marketing needs from the front-end users to the sales backend.

The development of Assistant does not stop at simple functionality implementation, but realizes efficient business processes through the careful design of workflows. Through the flexible combination of Q&A nodes, message nodes and workflows, the assistant is able to respond quickly according to user needs, improving user experience and sales conversion rate.

The main purpose of this post is to provide implementation ideas and how to better develop an assistant than just simply taking it apart. If you take the disassembly approach, an article may be as long as 20,000+ words, and also need to be accompanied by dozens of pictures, which will be very tedious. Therefore, for the details of the disassembly, I plan to create a separate video to help you understand more clearly. Thank you for your attention and support to Xiaoyu.


I'm Rain, a Java server-side coder, studying the mysteries of AI technology. I love technical communication and sharing, and I am passionate about open source community. I am also a Tencent Cloud Creative Star, Ali Cloud Expert Blogger, Huawei Cloud Enjoyment Expert, and Nuggets Excellent Author.

💡 I won't be shy about sharing my personal explorations and experiences on the path of technology, in the hope that I can bring some inspiration and help to your learning and growth.

🌟 Welcome to the effortless drizzle! 🌟