Tutorial: Semantic Clustering of Person Messages with LLM Prompts

As a Developer Advocate, it’s difficult to maintain up with person discussion board messages and perceive the large image of what customers are saying. There’s loads of priceless content material — however how are you going to rapidly spot the important thing conversations? On this tutorial, I’ll present you an AI hack to carry out semantic clustering just by prompting LLMs!

TL;DR 🔄 this weblog publish is about how one can go from (knowledge science + code) → (AI prompts + LLMs) for a similar outcomes — simply quicker and with much less effort! 🤖⚡. It’s organized as follows:

• Inspiration and Information Sources
• Exploring the Information with Dashboards
• LLM Prompting to supply KNN Clusters
• Experimenting with Customized Embeddings
• Clustering Throughout A number of Discord Servers

Inspiration and Information Sources

First, I’ll give props to the December 2024 paper Clio (Claude insights and observations), a privacy-preserving platform that makes use of AI assistants to research and floor aggregated utilization patterns throughout thousands and thousands of conversations. Studying this paper impressed me to do this.

Information. I used solely publicly accessible Discord messages, particularly “discussion board threads”, the place customers ask for tech assist. As well as, I aggregated and anonymized content material for this weblog.  Per thread, I formatted the info into dialog flip format, with person roles recognized as both “person”, asking the query or “assistant”, anybody answering the person’s preliminary query. I additionally added a easy, hard-coded binary sentiment rating (0 for “not glad” and 1 for “glad”) based mostly on whether or not the person stated thanks anytime of their thread. For vectorDB distributors I used Zilliz/Milvus, Chroma, and Qdrant.

Step one was to transform the info right into a pandas knowledge body. Beneath is an excerpt. You possibly can see for thread_id=2, a person solely requested 1 query. However for thread_id=3, a person requested 4 completely different questions in the identical thread (different 2 questions at farther down timestamps, not proven under).

I added a naive sentiment 0|1 scoring perform.

def calc_score(df):
   # Outline the goal phrases
   target_words = ["thanks", "thank you", "thx", "🙂", "😉", "👍"]


   # Helper perform to test if any goal phrase is within the concatenated message content material
   def contains_target_words(messages):
       concatenated_content = " ".be a part of(messages).decrease()
       return any(phrase in concatenated_content for phrase in target_words)


   # Group by 'thread_id' and calculate rating for every group
   thread_scores = (
       df[df['role_name'] == 'person']
       .groupby('thread_id')['message_content']
       .apply(lambda messages: int(contains_target_words(messages)))
   )
   # Map the calculated scores again to the unique DataFrame
   df['score'] = df['thread_id'].map(thread_scores)
   return df


...


if __name__ == "__main__":
  
   # Load parameters from YAML file
   config_path = "config.yaml"
   params = load_params(config_path)
   input_data_folder = params['input_data_folder']
   processed_data_dir = params['processed_data_dir']
   threads_data_file = os.path.be a part of(processed_data_dir, "thread_summary.csv")
  
   # Learn knowledge from Discord Discussion board JSON information right into a pandas df.
   clean_data_df = process_json_files(
       input_data_folder,
       processed_data_dir)
  
   # Calculate rating based mostly on particular phrases in message content material
   clean_data_df = calc_score(clean_data_df)


   # Generate reviews and plots
   plot_all_metrics(processed_data_dir)


   # Concat thread messages & save as CSV for prompting.
   thread_summary_df, avg_message_len, avg_message_len_user = 
   concat_thread_messages_df(clean_data_df, threads_data_file)
   assert thread_summary_df.form[0] == clean_data_df.thread_id.nunique()

Exploring the Information with Dashboards

From the processed knowledge above, I constructed conventional dashboards:

• Message Volumes: One-off peaks in distributors like Qdrant and Milvus (presumably as a consequence of advertising and marketing occasions).
• Person Engagement: Prime customers bar charts and scatterplots of response time vs. variety of person turns present that, on the whole, extra person turns imply larger satisfaction. However, satisfaction does NOT look correlated with response time. Scatterplot darkish dots appear random with regard to y-axis (response time). Perhaps customers should not in manufacturing, their questions should not very pressing? Outliers exist, reminiscent of Qdrant and Chroma, which can have bot-driven anomalies.
• Satisfaction Traits: Round 70% of customers seem glad to have any interplay. Information be aware: be certain to test emojis per vendor, typically customers reply utilizing emojis as an alternative of phrases! Instance Qdrant and Chroma.

LLM Prompting to supply KNN Clusters

For prompting, the subsequent step was to mixture knowledge by thread_id. For LLMs, you want the texts concatenated collectively. I separate out person messages from whole thread messages, to see if one or the opposite would produce higher clusters. I ended up utilizing simply person messages.

With a CSV file for prompting, you’re able to get began utilizing a LLM to do knowledge science!

!pip set up -q google.generativeai
import os
import google.generativeai as genai


# Get API key from native system
api_key=os.environ.get("GOOGLE_API_KEY")


# Configure API key
genai.configure(api_key=api_key)


# Record all of the mannequin names
for m in genai.list_models():
   if 'generateContent' in m.supported_generation_methods:
       print(m.title)


# Attempt completely different fashions and prompts
GEMINI_MODEL_FOR_SUMMARIES = "gemini-2.0-pro-exp-02-05"
mannequin = genai.GenerativeModel(GEMINI_MODEL_FOR_SUMMARIES)
# Mix the immediate and CSV knowledge.
full_input = immediate + "nnCSV Information:n" + csv_data
# Inference name to Gemini LLM
response = mannequin.generate_content(full_input)


# Save response.textual content as .json file...


# Test token counts and evaluate to mannequin restrict: 2 million tokens
print(response.usage_metadata)

Sadly Gemini API stored slicing brief the response.textual content. I had higher luck utilizing AI Studio instantly.

My 5 prompts to Gemini Flash & Professional (temperature set to 0) are under.

Immediate#1: Get thread Summaries:

Given this .csv file, per row, add 3 columns:
– thread_summary = 205 characters or much less abstract of the row’s column ‘message_content’
– user_thread_summary = 126 characters or much less abstract of the row’s column ‘message_content_user’
– thread_topic = 3–5 phrase tremendous high-level class
Be certain that the summaries seize the principle content material with out shedding an excessive amount of element. Make person thread summaries straight to the purpose, seize the principle content material with out shedding an excessive amount of element, skip the intro textual content. If a shorter abstract is sweet sufficient desire the shorter abstract. Be certain that the subject is normal sufficient that there are fewer than 20 high-level subjects for all the info. Desire fewer subjects. Output JSON columns: thread_id, thread_summary, user_thread_summary, thread_topic.

Immediate#2: Get cluster stats:

Given this CSV file of messages, use column=’user_thread_summary’ to carry out semantic clustering of all of the rows. Use method = Silhouette, with linkage methodology = ward, and distance_metric = Cosine Similarity. Simply give me the stats for the tactic Silhouette evaluation for now.

Immediate#3: Carry out preliminary clustering:

Given this CSV file of messages, use column=’user_thread_summary’ to carry out semantic clustering of all of the rows into N=6 clusters utilizing the Silhouette methodology. Use column=”thread_topic” to summarize every cluster subject in 1–3 phrases. Output JSON with columns: thread_id, level0_cluster_id, level0_cluster_topic.

Silhouette Rating measures how related an object is to its personal cluster (cohesion) versus different clusters (separation). Scores vary from -1 to 1. The next common silhouette rating usually signifies better-defined clusters with good separation. For extra particulars, try the scikit-learn silhouette rating documentation.

Making use of it to Chroma Information. Beneath, I present outcomes from Immediate#2, as a plot of silhouette scores. I selected N=6 clusters as a compromise between excessive rating and fewer clusters. Most LLMs as of late for knowledge evaluation take enter as CSV and output JSON.

From the plot above, you possibly can see we’re lastly entering into the meat of what customers are saying!

Immediate#4: Get hierarchical cluster stats:

Given this CSV file of messages, use the column=’thread_summary_user’ to carry out semantic clustering of all of the rows into Hierarchical Clustering (Agglomerative) with 2 ranges. Use Silhouette rating. What’s the optimum variety of subsequent Level0 and Level1 clusters? What number of threads per Level1 cluster? Simply give me the stats for now, we’ll do the precise clustering later.

Immediate#5: Carry out hierarchical clustering:

Settle for this clustering with 2-levels. Add cluster subjects that summarize textual content column “thread_topic”. Cluster subjects needs to be as brief as potential with out shedding an excessive amount of element within the cluster that means.
– Level0 cluster subjects ~1–3 phrases.
– Level1 cluster subjects ~2–5 phrases.
Output JSON with columns: thread_id, level0_cluster_id, level0_cluster_topic, level1_cluster_id, level1_cluster_topic.

I additionally prompted to generate Streamlit code to visualise the clusters (since I’m not a JS professional 😄). Outcomes for a similar Chroma knowledge are proven under.

I discovered this very insightful. For Chroma, clustering revealed that whereas customers have been pleased with subjects like Question, Distance, and Efficiency, they have been sad about areas reminiscent of Information, Shopper, and Deployment.

Experimenting with Customized Embeddings

I repeated the above clustering prompts, utilizing simply the numerical embedding (“user_embedding”) within the CSV as an alternative of the uncooked textual content summaries (“user_text”).I’ve defined embeddings intimately in earlier blogs earlier than, and the dangers of overfit fashions on leaderboards. OpenAI has dependable embeddings that are extraordinarily inexpensive by API name. Beneath is an instance code snippet how one can create embeddings.

from openai import OpenAI


EMBEDDING_MODEL = "text-embedding-3-small"
EMBEDDING_DIM = 512 # 512 or 1536 potential


# Initialize consumer with API key
openai_client = OpenAI(
   api_key=os.environ.get("OPENAI_API_KEY"),
)


# Perform to create embeddings
def get_embedding(textual content, embedding_model=EMBEDDING_MODEL,
                 embedding_dim=EMBEDDING_DIM):
   response = openai_client.embeddings.create(
       enter=textual content,
       mannequin=embedding_model,
       dimensions=embedding_dim
   )
   return response.knowledge[0].embedding


# Perform to name per pandas df row in .apply()
def generate_row_embeddings(row):
   return {
       'user_embedding': get_embedding(row['user_thread_summary']),
   }


# Generate embeddings utilizing pandas apply
embeddings_data = df.apply(generate_row_embeddings, axis=1)
# Add embeddings again into df as separate columns
df['user_embedding'] = embeddings_data.apply(lambda x: x['user_embedding'])
show(df.head())


# Save as CSV ...

Apparently, each Perplexity Professional and Gemini 2.0 Professional typically hallucinated cluster subjects (e.g., misclassifying a query about sluggish queries as “Private Matter”).

Conclusion: When performing NLP with prompts, let the LLM generate its personal embeddings — externally generated embeddings appear to confuse the mannequin.

Clustering Throughout A number of Discord Servers

Lastly, I broadened the evaluation to incorporate Discord messages from three completely different VectorDB distributors. The ensuing visualization highlighted frequent points — like each Milvus and Chroma dealing with authentication issues.

Abstract

Right here’s a abstract of the steps I adopted to carry out semantic clustering utilizing LLM prompts:

1. Extract Discord threads.
2. Format knowledge into dialog turns with roles (“person”, “assistant”).
3. Rating sentiment and save as CSV.
4. Immediate Google Gemini 2.0 flash for thread summaries.
5. Immediate Perplexity Professional or Gemini 2.0 Professional for clustering based mostly on thread summaries utilizing the identical CSV.
6. Immediate Perplexity Professional or Gemini 2.0 Professional to write down Streamlit code to visualise clusters (as a result of I’m not a JS professional 😆).

By following these steps, you possibly can rapidly rework uncooked discussion board knowledge into actionable insights — what used to take days of coding can now be executed in only one afternoon!

References

1. Clio: Privateness-Preserving Insights into Actual-World AI Use, https://arxiv.org/abs/2412.13678
2. Anthropic weblog about Clio, https://www.anthropic.com/analysis/clio
3. Milvus Discord Server, final accessed Feb 7, 2025
   Chroma Discord Server, final accessed Feb 7, 2025
   Qdrant Discord Server, final accessed Feb 7, 2025
4. Gemini fashions, https://ai.google.dev/gemini-api/docs/fashions/gemini
5. Weblog about Gemini 2.0 fashions, https://weblog.google/expertise/google-deepmind/gemini-model-updates-february-2025/
6. Scikit-learn Silhouette Rating
7. OpenAI Matryoshka embeddings
8. Streamlit