This article outlines tips and tricks around prompt engineering in AlphaSense's natural language interface, Generative Search.
What is prompt engineering?
Prompt engineering is the process of creating and refining "instructions" (example: prompts) for generative AI models to produce desired outputs. Put in the context of AlphaSense, it's the process of crafting the questions you'll ask our generative search tool to receive the information you're wanting.
Basic Prompt Engineering tactics
With natural-language search, you can leverage many of the same filters you would in a traditional search, like:
Timeframe:
By default, Gen Search will index the last 12 months. If you'd like to change the timeframe it searches against, use natural language in your question, such as
"...in the last 3 years"
"In Q2 of this year...",
"...last month..."
Companies:
You can search for companies by using the company name or ticker.
"How did Nvida perform..."
"...guidance from WFC?"
"What factors are affecting Stripe, Inc in...."
If searching for small private companies that have names shared with other brands (i.e. "The Standard" is the marketing name for Standard Insurance Company, but also the name of popular hotel chain, etc), we encourage you to add context or use the official company name to help out Generative Search in interpreting the question.
"What did Standard Insurance Company say..."
"...affect The Standard hotel chain?"
Industries:
We recommend to use words that speak directly to what you're looking for
Searching "electric vehicle industry" is more optimal compared to "electric vehicle space"
This removes uncertainty from the Generative Search algorithm and provides a more targeted answer
Keywords/Themes:
When possible, it's best to use keywords that will activate smart synonyms
"What R&D updates in the GenAI industry..."
"...the TAM of semiconductors?"
Watchlists:
Example Prompts for AlphaSense Content
Company Research
"What strategic initiatives is [COMPANY] implementing to address the profitability and growth issues it has experienced?"
"How are changes in regulatory environments and trade policies affecting [COMPANY]’s global operations and market strategies?"
"Give me a brief overview of [COMPANY]."
Industry Research
"What is the TAM for [INDUSTRY] in North America?"
"What are the key factors driving capital investment in [INDUSTRY] over the next 5 to 10 years?"
"What do industry experts say about major factors affecting the [INDUSTRY] industry?"
Comparison
"Compare the broker view of [COMPANY]'s competitive positioning with expert views."
"Compare and contrast bull and bear views on [COMPANY/TOPIC]."
"Compare R&D pipelines of [COMPANY 1] vs. [COMPANY 2]"
Event Summarization
"Summarize the expert view on [COMPANY]'s growth potential over the next 5 years."
"Summarize the broker view of [COMPANY]."
"Summarize [COMPANY]'s latest investor presentation."
Change Over Time
"How have broker forecasts for [COMPANY] earnings changed?"
"How have expert opinions on opportunities and challenges for the [INDUSTRY] sector evolved?
Searching Specific Content
Content Types and Example Language
Content in AlphaSense can be referred to in a variety of ways. The below dictionary provides a few examples for each
Content Type | Example Language |
Company Documents | "What is <Company> guidance..." "...as per 10K" "How does the <Company> plan..." |
Broker Research | "How do brokers view..." "What are analysts saying..." "What does <Broker> say about..." |
Expert Insights | "What's the expert position on..." "What are industry experts saying..." "What do former competitors believe..." |
News & Regulatory | "What are news agencies saying..." "List the latest news on..." "...within PubMed articles" |
Internal Content** | "Do we have any internal data about..." "What is our internal perspective on..." "What are our views on..."
see below for more on internal content |
** If you're an Enterprise Intelligence user of AlphaSense, you'll also be able to leverage Generative Search across your integrated, internal content in addition to the content existing in AlphaSense
What content does Generative Search index by default?
There are 2 different avenues available for you to access and launch Generative Search:
Directly launch Gen Search from the left hand navigation panel
Dive into Gen Search after you've run a traditional Document search
When you begin your search with Generative Search [Option 1], it accesses your default sources with no content filters applied. This allows you to search wide and use natural language to filter for various sources and content.
If you launched Generative Search after running a Document search [Option 2], Gen Search will carry over any content filters applied to those searches into your chat. If you need to clear any filters, either, select "Clear Chat" or start a new session with Gen Search.
Example: A search is run for "R&D strategy" and "PFE", as well as a content filter for Expert Insights is applied
If you open Gen Search from this query, the summaries will only be generated from Expert Insights content.
Searching Internal Content
If you're an Enterprise Intelligence customer, AlphaSense Generative Search is extended to your internal content within AlphaSense.
Using Pronouns
Use pronouns instead of your company name to ask questions specific to your internal content:
"What are the semiconductor investment targets that we identified?"
"Synthesize our internal research on edge computing trends."
Searching Colleagues
You can prompt what specific authors have said about a particular topic to isolate that particular voice. Make sure to use both First and Last Name.:
"What has [COLLEAGUE'S NAME] said about EV growth levers in our internal content?"
Internal Content Filters
Using terms like "our", "my", and "we" will direct generative search to filter content specifically to any internal content you've added to AlphaSense
Using the specific source in your prompt will filter for that particular content set
Example: "In our Uploads...", "In FileSync...", "Our content in Box..."
Example Prompts for Internal Content
Company Research
"According to our research, what are [COMPANY]'s key growth drivers?"
"Summarize our investment thesis for [COMPANY]."
"Summarize the motivations behind [COMPANY]'s acquisition of [COMPANY] according to our research."
Industry Research
"Create a market landscape of the [INDUSTRY] based on our internal content."
"Synthesize our internal research on [INDUSTRY]."
"What has the historical growth trajectory of [INDUSTRY] looked like per our internal documents?"
Comparisons
"Compare our view of [COMPANY]'s competitive positioning with external views."
"Evaluate our outlook on [COMPANY] against broker consensus."
"Compare our view on [INDUSTRY] growth drivers against the broker view."
Event Summarization
"Summarize key points from our management meeting with [COMPANY]."
"What were the key points discussed at [COMPANY]'s latest investor presentation according to our internal research?
Change Over Time
"How has our team's perspective on [COMPANY] changed overtime?"
"What has [COMPANY]'s revenue growth looked like over the past 5 years according to our data?"
Advanced Prompt Engineering Techniques
Conversational Memory
Interactions with our generative search are conversational, meaning that it will "remember" previous queries from that session and, by default, assume that your question is related to previous queries.
You can start a new "conversation" by clicking the "Clear Chat" button in the upper right hand corner of the screen.
Requesting Quantitative Information
Adding "use numbers" or “data” leads to a more quantitative answer from Gen Search.
"What is predicted AI investment in the industrial sector? Include numbers."
"What are some important growth drivers for semiconductors? Use data"
Driving More Direct Responses
We recommend prefacing a prompt with “explain” yields a more direct response than more ambiguous prompts.
"What is predicted AI investment in the industrial sector? Explain why."
"Explain how Netflix guidance..."
Optimizing Comparison Questions
To get the best answers from comparison questions, we recommend specifying a perspective/document type in your prompt
"How are analysts talking about Netflix's growth potential vs. Hulu"
"Compare broker's competitive positioning of PFE against former employees."
Explore Suggested Questions Following a Response
Oftentimes, exploring the suggested follow up questions in Gen Search can help reframe the question to be more specific or direct in a way that is going to warrant a better response from the model.
These recommendations are really great for idea generation as well
Review the Summary/Plan
Reviewing the Summary provides insight insight into your prompt and if there are any immediate adjustments needed. Checking if the summary is showing the correct timeframe, content, keywords, etc can verify the prompt functioned as you intended it to.
Current unsupported prompts with Generative Search
Watchlists:
There is NOT currently a way for you to ask Generative Search about a watchlist. For the time being, we suggest that you leverage traditional document search to query about topics over your watchlists, and then you can explore the search summary and ask follow-up questions from there.
Saving conversations:
At this time, there is no mechanism for saving conversation threads with Generative Search to revisit at a later time.
Librarian questions:
You might want to ask Gen Search to bring relevant documents rather than respond to a question (i.e. "Bring me the last 3 earnings transcripts for GM"). Rather than using Gen Search, you should perform a traditional Document Search (i.e. company = GM, filter for earnings transcripts).
Screener questions (aka producing lists):
Generative Search will likely give a decent overview to a request like "Who are the top 5 players in biologics by market cap?", but a more exhaustive, data-driven list can be found with Companies in Search Results after completing a traditional document search.
Ask across specific document(s):
You may want to define the universe of documents that Gen Search is summarizing very specifically, even limiting it to be only one document. At this time, generative search is designed to synthesize large swaths of documents, rather than one or a few.
Surfacing tables/charts:
For the time being, Image Search (after running a traditional document search) should be the way you surface tables and charts about a topic/area of interest.