Generate Email Replies with a Semantic Search Memory

Build a retrieval-augmented LLM workflow that summarizes incoming emails, drafts replies, maintaining a semantic memory of past messages

Responding to repetitive business emails often requires recalling similar past messages and constraints such as scheduling conflicts. This example demonstrates how to store short email summaries in a semantic index and use them to generate context-aware reply drafts.

Define a sample incoming email that will serve as the first message:

In[1]:=

email="Dear Dr. Smith,We are honored to have you as a keynote speaker at the Future of Technology Summit, scheduled from June 11 to June 13. Your insights on the future of technology will be invaluable to our attendees.Kind regards,Summit Coordinator";

Create a semantic search index that stores each email together with an LLM-generated one-sentence summary:

In[2]:=

summary=LLMFunction["Read the following email and provide a one sentence summary: ``"][email]

Out[2]=

Dr. Smith is invited to be a keynote speaker at the Future of Technology Summit from June 11 to June 13.

Create a simple semantic search index with only one entry:

In[3]:=

CreateSemanticSearchIndex[summaryemail,"emails"]

Out[3]=

SemanticSearchIndex

ID: emails

Elements: 1



Define a second email that will be processed using the retrieval-augmented workflow:

In[4]:=

newemail="Dear Dr. Smith,We are pleased to invite you to speak at the AI and Innovation Conference, which will take place from June 10 to June 12. Your expertise in computational thinking and artificial intelligence would greatly enrich our event.Best regards,Conference Organizer";

Create a graph that (1) summarizes the new email, (2) drafts a reply using prompts grounded in the semantic index, and (3) appends the new summary-email pair back into the index:

In[5]:=

result=LLMGraph["Summarize""Read the following email and provide a one sentence summary: `Email`","Answer"LLMFunction["propose an answer to this email `Email`",LLMEvaluator"Prompts"LLMPromptGenerator[SemanticSearchIndex["emails"]]],"Append"Function[UpdateSemanticSearchIndex["emails",#Summarize#Email]]][newemail,"LLMGraph"]

Out[5]=

LLMGraph

Input:

Output:

Answer

Append



Extract the generated summary:

In[6]:=

Information[result,"Nodes"]["Summarize","Result"]

Out[6]=

Dr. Smith is invited to speak at the AI and Innovation Conference from June 10 to June 12 due to his expertise in computational thinking and artificial intelligence.

Inspect the proposed reply;

In[7]:=

Information[result,"Nodes"]["Answer","Result"]

Out[7]=

Dear Conference Organizer,Thank you very much for the kind invitation to speak at the AI and Innovation Conference. I truly appreciate the opportunity to share my insights on computational thinking and artificial intelligence.However, I regret to inform you that I am already committed to speaking at the Future of Technology Summit from June 11 to June 13. Unfortunately, this conflict prevents me from participating in your event on June 10 to June 12.I wish you great success with the conference, and I hope to have the opportunity to collaborate in the future.Best regards, Dr. Smith

Publisher Information

Contributed by: Wolfram Staff

License Information

This work is licensed under a Creative Commons Attribution 4.0 International License

Wolfram Language Example Repository

Generate Email Replies with a Semantic Search Memory

See Also

Publisher Information

License Information

Generate Email Replies with a Semantic Search Memory

See Also

Related Symbols

Publisher Information

License Information