How does the AI decide what needs a response?
Our AI uses natural language processing (NLP) to understand the user's intent to receive a response, or in other words, it finds out if the user is looking for a reply from your brand. The AI is trained to detect various types of conversational intent.
It walks through how messages are flagged for response, how prioritization works within the Smart Inbox, and how your team can take action based on AI recommendations.
Which AI models do you use?
We use a combination of our own proprietary AI model, trained over 11+ years on billions of social media conversations, and trusted third-party tools for specific use cases.
Our in-house model is built specifically for social, capable of understanding slang, sentiment, emojis, and fast-moving trends with high accuracy.
We also integrate third-party models for specialized tasks:
OpenAI & Gemini: Power one-click AI-generated replies, captions, and chat-based analysis through features like Astra the Analyst
Translation models: Enable real-time multilingual moderation
What is sent to Astra?
The data Astra receives depends on which tab or workflow you're using it from within the BrandBastion platform.
In Analyze > Listen
Astra receives:
All data from the dashboard:
Overview
Sentiment Details
Topics
Up to 10,000 recent comments, processed in batches
What’s excluded:
Private messages
Comments with only user tags
Username-level data
Messages tagged as PII or spam (filtered using BrandBastion’s advanced tagging)
YouTube data
This process ensures performance and relevance at scale:
It leverages pre-analyzed dashboard insights plus comment batches.
Astra processes comments in structured batches and scales up to the model’s input limit, so analysis reflects a much broader dataset while staying efficient.
For instance, sentiment distribution, platform breakdowns (e.g. Facebook vs. Instagram), and top tags still reflect the full dataset, not just the comments included in any single batch.
This approach avoids inefficient raw analysis and relies on BrandBastion's structured tagging system to guide Astra’s topic identification.
Why are comments processed in batches?
Astra works best when the input is structured and within model limits. Batch processing lets Astra analyze up to 10,000 comments while keeping insights consistent, fast, and grounded in the same dashboard context.
In Posts tab
Astra receives data from the most recent 1,000 posts, sorted by publish date. For each post, Astra receives:
Platform
Post copy
Date
Format (image, video, etc.)
Total views and engagement metrics (reactions, comments, shares)
Engagement rate (ER) and Net Sentiment Score (NSS)
Comment breakdown (positive, negative, neutral, hidden, brand, tagged)
Post type and tags
What’s excluded:
YouTube posts
Posts without view data
This structured post-level input helps Astra:
Compare engagement across platforms and formats
Understand comment sentiment trends at post level
Surface insights tied to specific post tags (e.g. campaign or topic-level analysis)
In Competitor Analysis tab
Astra receives up to 4,000 reviews from your own apps and up to 4,000 reviews from each competitor. For each review, Astra receives:
Your brand reviews:
Rating
Review text
Country
Platform
App name and version
Review date
Sentiment
Custom tags
Brand reply (if one exists)
Competitor reviews:
Rating
Review title
Review text
Country
Platform
App name and version
Review date
Language
Helpfulness votes
In Reviews tab
Astra receives up to 10,000 reviews, processed in batches. For each review, Astra receives:
Rating (stars, or Recommends / Doesn't Recommend for Facebook)
Review text
Country and platform
App name and version
Review date, including whether the review was edited and how the rating changed
Sentiment and custom tags
Brand reply (if one exists)
Up to 2 previous versions of the review (edit history)
Astra also receives aggregated dashboard data for the same time period:
Review volume
Average rating
Rating distribution
Platform comparison
Ratings by country
Ratings by app version
What’s excluded:
YouTube
Brand replies that were ignored or have no content