AccessAI™ Release Notes
October 27, 2025

Resources

• Large Language Model (LLM) Overview
• Examples Of What LLM’s Are Good At
• Examples Of What LLM’s Cannot Do
• Prompting AccessAI
• Bug Fixes

Large Language Model (LLM) Overview

• AccessAI is a complex system composed of a knowledge collection, Large Language Model (LLM), and algorithms for processing prompts.
• An LLM sits at the core of the system and give it the ability to understand natural language and generate responses.
• LLMs are nearly magical, but they have some limitations too. Understanding what LLMs are good at and bad at can help you use AccessAI™ most effectively.

Examples Of What LLM’s Are Good At

1. Text generation/transformation

LLMs tend to be good at producing lots of language. So if you ask an LLM to generate a couple of paragraphs explaining and providing background for a topic, you’re likely to get something pretty usable.

LLMs are also very good at transforming language between formats. For example, maybe you want your explanation of a new piece of legislation formatted for social media posting. Remember that LLMs have consumed most every type of language on the public internet (from human languages to programming languages), from tweets to blog posts to scientific articles, and can emulate most any style that they’ve been exposed to.

2. Brainstorming

As a kind of subset of text generation, LLMs are pretty good at brainstorming ideas. Core to the brainstorming process is generating a lot of ideas and LLMs tend to do that very well. Asking for 10 ideas will likely give you a decent list, but if you need 50 or 100 or more, LLMs are readily able to provide you with those.

The solution or idea you end up with won’t be entirely the LLM’s creation, and you’ll definitely have to provide some coaching and bring in some of your own knowledge—but as a pure idea generation machine, large language models are pretty unmatched.

3. Summarizing

Because of how they were created and trained, LLMs are great at summarization. The training process of an LLM involves compressing information at its core, distilling lengthy text into more concise forms without losing essential meaning, which is exactly what you’re doing when creating a summary. In addition, the transformers that LLMs are largely based on, are particularly good at understanding context and relationships in the text. Attention mechanisms within transformers allow the model to focus on relevant parts of the input when generating summaries.

Did you know: transformers are the “T” in “GPT”?

One thing you’ll want to keep in mind is, because of how LLMs work, you can’t really ask for the summary to be a specific word count. However, you can continue to ask for shorter, longer, simpler, or more complex summaries until you get one that works for you.

Examples Of LLM Limitations

1. Hallucinations (Making Up Information)

One weird thing about LLMs is that when they don’t know the answer, they often won’t admit it. Instead, they’ll confidently make up something that sounds believable. This is called a “hallucination.” For example, if you ask for a fact about a historical event that wasn’t in the data it was trained on, the LLM might invent details or events that never happened.

2. Limited Reasoning Skills

Even though LLMs can seem very smart, they often struggle with basic math. This is because they weren’t really designed to solve math problems. While LLMs are good at understanding and generating sentences, they’re not great at solving complex problems. For example, if you ask an LLM to solve a multi-step math problem or a puzzle, it might get confused and make mistakes along the way.

3. Limited Long-Term Memory

Each time you use an LLM, it starts with a blank slate—it doesn’t remember your previous conversations unless you remind it in the current session. This can be frustrating if you’re trying to have an ongoing discussion or work on a project over time.

4. Limited Knowledge

LLMs are trained on data from the past. It means that if LLMs don’t have access to the internet or any way to look up information in real time, they don’t know anything that happened after their training data was collected. If you ask about recent events, they won’t be able to provide accurate answers.

5. Bias

LLMs learn from the text they’re trained on, and that text comes from the internet or, in the case of a Directed Large Language Model, from you.  As a result, LLMs can sometimes reflect the same biases in their responses that you have.

Features within AccessAI™

1. Spreadsheet Optimized

This functionality optimizes analysis of ingested spreadsheets (e.g. Excel, CSV) to maximize synthesis and interpretation of tabular data.

Key features at a glance:

• Cell-aware reasoning: The model treats inputs and outputs as ranges, cells, and formulas, enabling reliable editing and creation of sheets without breaking references.
• Deterministic transformations: For repetitive tasks—cleaning data, normalizing formats, or generating formulas—the assistant applies consistent rules to ensure predictable results.
• Schema adherence: It respects column types and headers, surfacing validation errors and proposing fixes instead of silently coercing data.
• Formula fluency: Deep understanding of spreadsheet functions (e.g., VLOOKUP/XLOOKUP, INDEX/MATCH, SUMIFS, QUERY) and array formulas, including cross-sheet references and named ranges.
• Minimal hallucination: When a formula or dataset is ambiguous, the mode prompts for the smallest possible clarification to avoid incorrect outputs.
• Export-ready outputs: Produces CSV/TSV blocks, formula lists, and step-by-step transformations that paste cleanly into spreadsheet software.

How it works under the hood:

• Structured planning: The assistant maps user goals to a sequence of spreadsheet operations (insert columns, add validations, compute derived fields).
• Constraint checking: Before emitting results, it simulates formula evaluation and checks for common errors (#REF!, #N/A, circular refs) and range mismatches.
• Data-first formatting: Outputs are aligned to the requested schema, with explicit headers and consistent value types (dates, numbers, text).

What users experience:

• Clear prompts like “Paste your headers and 5 sample rows” to capture structure.
• Direct, pasteable results: “Add this formula to D2 and fill down.”
• Optional automations: The assistant can generate Apps Script snippets or explain how to apply steps in Excel/Google Sheets.

In short, “Spreadsheet Optimized” is a mode that biases the assistant toward reproducible, low-error spreadsheet work—prioritizing formula correctness, schema fidelity, and paste-ready outputs for real-world sheets.

2. Lookup

How it works at a glance:

• Triggering: The model detects a gap in its knowledge or the need for precision (dates, figures, code APIs). That recognition triggers a retrieval step.
• Retrieval: The system queries a designated source—this could be a search API, a database, a knowledge base, or tool-specific endpoint—returning relevant snippets.
• Grounding: The fetched content is parsed, filtered, and ranked for relevance and credibility.
• Synthesis: The model integrates the grounded facts with its generative reasoning to produce a final response, often citing or linking the sources depending on configuration.
• Caching and reuse: Results may be cached temporarily to reduce redundant lookups and improve latency for similar queries.

Common benefits and safeguards:

• Freshness: Access to current data beyond the model’s training cutoff.
• Accuracy: Reduces hallucinations by grounding answers in retrieved text.
• Traceability: Enables source attribution and auditing in some setups.
• Safety & policy filters: Retrieved content is passed through moderation and policy checks before being shown.

In short, “Lookup” augments the model with retrieval so it can fact-check, stay current, and answer with greater precision, while balancing latency and reliability through caching and filters.

3. Text Compare

The operation of a “Text Compare” feature in a Language Model (LLM) generally involves several key steps, even though specific details might vary depending on the implementation. Here is a general workflow:

• Preprocessing: The text inputs are first preprocessed. This step may include cleaning the text by removing unnecessary characters, normalizing the text to a consistent format, and possibly segmenting the text into smaller units if needed.
• Representation (Tokens/Embeddings): The preprocessed text is then converted into a numerical format that the LLM can understand. This is typically done by tokenizing the text into smaller units (tokens) and then converting these tokens into embeddings, which are dense vector representations of the text.
• Similarity or Diff Computation: Once the text is represented as embeddings, the LLM can compute the similarity or differences between the texts. This is often done using mathematical operations such as cosine similarity, Euclidean distance, or other metrics that quantify how similar or different the embeddings are.
• Output Formatting: Finally, the results of the comparison are formatted into a human-readable output. This might include highlighting differences, providing a similarity score, or generating a summary of the comparison results.

This workflow allows the LLM to effectively compare texts by leveraging its understanding of language through embeddings and similarity computations.

4. Chat with Cabinet

Chat with Cabinet allows the user to direct the LLM to reference specific files for answering a prompt, thus limiting the answer to a question to files that have likely been pre-vetted as an “expert” source for a response.  This “one source of the truth” feature is important when expert resources are prioritized above other third party sources.

5. Content Management

AccessAI™ can ingest files of any type (PDF, HTML, PowerPoint, Excel, csv, etc.) and has the ability to extract files from online urls.  AccessAI™ is a “Directed” LLM in that each user can decide which content they want to select to add to the foundational content provided by the Pharmspective team to AccessAI™.  This “user-directed” experience allows organizations or individuals to avoid content posted on the internet that has not been pre-vetted for accuracy.

The ingestion portal also allows for the viewing of files that failed ingestion for troubleshooting investigation.