Bug Fixes

Summary

This bug fix addresses the issue where users were unable to see the upload progress for XLS files in thread or chat environments. Previously, when users uploaded XLS files, there was no indication of progress, leading to confusion and uncertainty about whether the upload was successful. This fix introduces a progress indicator, enhancing the user experience by providing real-time feedback during the upload process.

Technical Details

To implement this fix, we integrated a task and watcher system specifically for XLS file uploads. This involves:

• Task Initialization: A task is created when an XLS file upload is initiated. This task is responsible for managing the upload process and tracking its progress.
• Watcher Integration: A watcher is set up to monitor the task’s status. It periodically checks the progress of the upload and updates the user interface accordingly.
• Validation Steps: We conducted thorough validation on different environments:
  • Development Environment: Initial testing to ensure the task and watcher are functioning as expected.
  • Staging Environment: Further testing to simulate real-world conditions and ensure stability before deployment.

Explain it to Me Like I’m Five

Imagine you’re building a LEGO tower, and you want to know how much you’ve built so far. Before, you couldn’t tell how tall your tower was until you finished. Now, with this fix, you have a ruler next to your tower that shows you how tall it is as you build. This way, you always know how much more you need to build to finish your tower.

Specific Items Shipped

• Task for XLS Uploads: We created a new task that starts whenever an XLS file is uploaded. This task is like a manager that keeps track of everything happening during the upload.
• Watcher for Progress Monitoring: A watcher was added to keep an eye on the upload task. It checks how far along the upload is and updates the screen to show this progress.

The Problem We Solved with This Fix

The main problem was the lack of feedback during XLS file uploads in threads or chats. Users were left guessing whether their files were uploading correctly, leading to frustration and potential errors. By adding a progress indicator, we provide clear feedback, improving user confidence and satisfaction.

Who We Built This For

This fix was primarily built for users who frequently upload XLS files in chat or thread environments. These users rely on timely and accurate feedback to ensure their files are uploaded successfully, which is crucial for maintaining smooth communication and workflow.

Summary

We addressed a critical issue where MP4 files were not functioning correctly in the production environment of Storytell. This malfunction was caused by a recent refactor that moved processing to a job-based system without properly outputting sanitized HTML. The fix ensures that MP4 files are now processed seamlessly, restoring their functionality across all environments.

Technical Details

The root cause of the MP4 failure was identified during the transition to a job-based processing system. In the refactor, the process responsible for handling MP4 files did not include the necessary step to output sanitized HTML, leading to failures in production. To resolve this, we:

• Reintroduced Sanitized HTML Output: Ensured that the job responsible for processing MP4 files correctly outputs sanitized HTML, preventing any security vulnerabilities and ensuring proper functionality.
• Updated Job Configuration: Modified the job settings in the Storytell AI Platform repository to handle MP4 files appropriately.
• Enhanced Validation Processes: Implemented thorough validation checks across development, staging, and production environments to ensure that MP4 processing works flawlessly and to catch any potential issues early in the deployment pipeline.

Explain It to Me Like I’m Five

Imagine you’re trying to show a video to your friends, but every time you try, nothing appears. We found out that when we changed how we prepare the videos, we forgot to include a special step that makes them show up correctly. Now, we’ve added that step back, so your videos play without any problems, just like they should.

Specific Items Shipped

• Sanitized HTML Output Implementation: Added the necessary code to ensure that sanitized HTML is correctly generated during MP4 processing jobs.
• Job Refactor Fix: Updated the job structure to handle MP4 files without causing failures in production.
• Comprehensive Validation: Conducted extensive testing in development, staging, and production environments to confirm the effectiveness of the fix and ensure no residual issues remain.

The Problem We Solved with This Feature

We built this feature to address the critical issue of MP4 files failing to work in the production environment of Storytell. This problem prevented users from uploading and viewing MP4 content, disrupting their experience and the platform’s functionality. By resolving this, we ensure that media-rich content can be shared and enjoyed seamlessly, maintaining user satisfaction and platform reliability.

Who We Built This For

This fix was specifically designed for Storytell users who rely on uploading and sharing MP4 videos as part of their storytelling experience. Whether it’s content creators sharing educational videos or users uploading media for personal projects, ensuring MP4 functionality is essential for their use cases. By resolving this issue, we support a smooth and uninterrupted experience for all users engaging with video content on our platform.

Summary

We have addressed this issue that was occurring in our production environment. This fix ensures that our tokenization strategy aligns with the models in use, preventing token overflow issues. Additionally, several optimizations and improvements have been implemented to enhance overall system performance.

Technical details

The error was caused by a mismatch in tokenization strategies between our platform and the deployed models. We were using the Cl100kBase tokenization strategy, while the 4o models utilize O200kBase. This discrepancy led to an excessive number of tokens being processed, resulting in errors in production. By updating our tokenization strategy to O200kBase, we ensure compatibility with the 4o models, thereby preventing the “Too Many Tokens” error. Furthermore, during the troubleshooting process, additional optimizations were made to improve system efficiency and reduce the likelihood of similar issues in the future.

Explain it to me like I’m five

Think of our system like a backpack that can only hold a certain number of books. We realized that we were trying to put too many books in the backpack, causing it to break. We fixed this by adjusting how many books we put inside, making sure the backpack stays strong and works well. We also made some other small improvements to keep everything running smoothly!

Specific items shipped

• Aligned Tokenization Strategy: Changed the tokenization method from Cl100kBase to O200kBase to match the 4o models, preventing token overflow errors.

• Performance Optimizations: Implemented various optimizations to enhance system performance and ensure smoother operations.

• Bug Fix Deployment: Successfully deployed the fix to the production environment, resolving the “Too Many Tokens” error.

The problem we solved with this fix

We built this fix to resolve the “Too Many Tokens” error that was disrupting our production environment. This error was caused by using an incompatible tokenization strategy, leading to an excessive number of tokens being processed and causing system failures. Addressing this issue is crucial to maintain the reliability and stability of our platform, ensuring a seamless experience for our users without interruptions or errors.

Who we built this for

This fix is designed for our engineering team and our end-users who depend on the reliability of our platform. By eliminating the token overflow issue, developers can continue to build and deploy applications without encountering the “Too Many Tokens” error, while end-users benefit from a more stable and efficient system experience.

Summary

The latest fix disables the YouTube scraping functionality temporarily while we undergo debugging to ensure the overall stability and performance of our system. This measure is in place to prevent any potential issues arising from incorrect data retrieval or processing during the debugging phase. By disabling this feature, we can focus on identifying and resolving critical bugs without the added complexity of YouTube data scraping.

Who we built this for

This fix is specifically aimed at developers and QA engineers within our team who are involved in the debugging process. By temporarily disabling YouTube scraping, we provide these users with a clearer environment, eliminating potential discrepancies caused by YouTube data during testing. This allows for a more straightforward debugging process, enhancing our ability to resolve issues effectively.

Summary

We’ve removed the “fallback” behavior from the CSV processing workflow. Previously, if embeddings couldn’t be generated for a CSV, the job would continue using a “best effort” approach. With this fix, if embeddings generation fails, the entire job will fail, clearly signaling to the user that something went wrong.

Technical Details

In the previous implementation, the CSV row processing included a fallback mechanism that attempted to proceed even when embeddings couldn’t be generated. This approach was intended to monitor real-world performance and necessitated sending additional announcements when failures occurred. However, due to the inherent unpredictability of using a Large Language Model (LLM) for template generation, we encountered a consistent failure rate of approximately 12% in production.

By removing the fallback behavior, the system now enforces a strict policy where the job fails if embeddings aren’t successfully generated. This change not only ensures immediate feedback to users about processing issues but also allows the embedding process to be retried a defined number of times (N attempts). This retry logic enhances reliability without compromising on clear communication of failures.

Explain it to me like I’m five

Imagine you’re building a LEGO model, and sometimes some pieces are missing. Before, if a piece was missing, you’d just keep building and hope for the best. Now, if a piece is missing, the whole project stops so you know there’s a problem. This way, you can fix the issue right away instead of ending up with an incomplete model.

Specific Items Shipped

• Removal of Fallback Mechanism: Eliminated the “best effort” approach in CSV processing to ensure that jobs fail when embeddings cannot be generated.

• Job Failure on Embedding Generation Failure: Implemented a system where the entire job fails if embeddings aren’t successfully created, providing immediate feedback to users.

• Retry Logic for Embedding Process: Added functionality to attempt the embedding process multiple times (N retries) before ultimately failing, increasing the chances of successful embeddings without hiding failures.

The Problem We Solved with This Fix

Previously, the CSV processing system would continue running even when embeddings generation failed, which led to incomplete or inaccurate data processing. This “best effort” approach masked underlying issues and did not effectively inform users about failures, resulting in a 12% failure rate in production. By enforcing a job failure when embeddings can’t be generated, we ensure that users are immediately aware of problems, allowing for quicker troubleshooting and maintaining the integrity of the data processing workflow.

Who We Built This For

We built this fix for data engineers and analysts who depend on reliable embeddings generation for processing CSV files. Use cases include preparing data for machine learning models, where accurate embeddings are crucial for model performance. By ensuring that embedding failures are promptly reported, we help these users maintain high data quality and streamline their workflow by addressing issues as they arise.

Summary

The “Fix CSV Sampling” update addresses an issue where the sample CSV file was incorrectly generating a single row that combined both the header and the actual data rows. This fix ensures that the CSV file is properly formatted with distinct headers and multiple data rows, allowing for accurate and efficient data handling.

Technical details

The core issue was identified in the CSV sampling module, where the function responsible for generating the sample was inadvertently appending data rows directly to the header row, resulting in a malformed single-row CSV file. To resolve this, the function has been refactored to separate the header generation from the data row accumulation. Specifically, the header is now initialized independently, and each subsequent data row is appended as a new distinct row in the CSV structure. Additionally, error handling has been enhanced to ensure that any discrepancies in data formatting are caught and addressed before file generation. This ensures that the resulting CSV adheres to standard formatting conventions, facilitating seamless integration with data processing tools and workflows.

Explain it to me like I’m five

Imagine you’re making a list of your favorite toys. First, you write the titles like “Toy Name” and “Color” at the top. Then, you add each toy’s details on different lines below. Before this fix, sometimes all the toys and the titles got mixed up on one line, making the list hard to read. Now, each toy has its own line under the correct titles, making the list neat and easy to understand.

Specific items shipped:

• Separated Header and Data Rows: The CSV generator now clearly distinguishes between the header row and the data rows, ensuring that each section is properly formatted and organized.

• Enhanced Error Handling: Improved mechanisms are in place to detect and handle any formatting issues during CSV creation, reducing the likelihood of errors in the generated files.

• Optimized Data Processing: Adjustments to the data processing logic allow for more efficient handling of large datasets, resulting in quicker and more reliable CSV file generation.

The problem we solved with this fix

Previously, the CSV sampling feature was producing files where the header and data were merged into a single row. This made the CSV files difficult to use with standard data tools, as they expected distinct headers and multiple data rows. By correcting this structure, we ensure that users can seamlessly import and manipulate their data without encountering formatting issues, thereby improving overall data reliability and usability.

Who we built this for

We developed this fix for data analysts and developers who rely on accurate CSV files for data manipulation and reporting. By ensuring properly formatted CSV samples, these users can efficiently import data into their preferred tools without encountering structural issues, thereby streamlining their workflow and enhancing productivity.

Summary

This fix introduces a more lenient approach to matching header rows when processing CSV files. The previous implementation required a strict mapping of every column, which did not accommodate cases where columns might be omitted by the LLM (Large Language Model) due to lack of value. The goal is to adapt our processing to better align with LLM behavior, specifically allowing it to choose relevant headers while ignoring unnecessary ones.

Technical details

The previous methodology enforced a rule whereby all columns in the incoming CSV file had to be accounted for during processing. This rigidity was problematic, especially when working with responses from LLMs like Claude, which often discard columns deemed irrelevant.

With this fix, the processing logic has been modified to accept LLM responses without requiring strict adherence to our original column mapping rules. Here’s how it works:

• We now evaluate the header rows with greater leniency and allow the LLM to determine which fields to include for processing.
• The system captures headers that contain actual data and ignores those like “Rk” (row number), which Claude frequently drops, improving both performance and usability.
• This shift not only aligns our software more closely with LLM behavior but also streamlines data handling, resulting in fewer errors during the CSV upload process.

Explain it to me like I’m five

Imagine you have a box of crayons, but sometimes some of your crayons don’t work or are just plain silly colors and you don’t want to use them. This fix helps our program listen to a friend (the computer) who says, “Hey, these crayons are the best ones. Let’s just use these!” Even if you have a set of rules about which crayons you must use, it’s better to listen to your friend and pick the pretty ones instead of sticking to old rules that don’t make sense. This way, we end up with a nicer picture!

Specific items shipped:

• Lenient Column Matching: The system now allows missing columns, which means if Claude decides certain columns aren’t needed, the process continues smoothly without errors.
• Support for Irrelevant Headers: Headers like “Rk” will be ignored if they yield no meaningful data. This optimizes the CSV processing by focusing only on important information.
• Enhanced Validation Process: New validations ensure that users’ CSVs work seamlessly under the new rules.

The problem we solved with this fix

The initial CSV processing system was too strict, leading to failures or unnecessary complexity in handling files that LLMs might process differently. This rigidity was a barrier for effective data manipulation and integration, causing delays and frustration. By adopting a more flexible approach, we are now able to better serve our users and allow the system to function more intuitively in line with LLM capabilities.

Who we built this for

We primarily built this feature for users who handle CSV data extensively, such as:

• Data Analysts: They require efficient data imports without errors to analyze and draw insights.
• Developers: Those interfacing with LLMs are often faced with CSV data that doesn’t conform strictly to preconceived rules, and this fix helps solidify that interaction.
• End Users: Users who upload various data records and need reliable processing without inflexibility.

Summary

We’ve implemented a fix that sets a maximum number of attempts for updating the status of an asset. This enhancement ensures that the system doesn’t get stuck trying indefinitely to update an asset’s status, which was causing errors and job failures previously. By limiting the number of attempts, we improve the reliability and stability of the asset processing workflow.

Technical details

The fix involves modifying the fileproc module to include a max_attempts parameter when updating an asset’s status. Previously, the system would continuously attempt to update the asset status without a defined limit, leading to repeated failures and resource exhaustion.

In the codebase, we’ve introduced a retry mechanism that caps the number of update attempts to a predefined maximum. If the asset status update fails, the system will retry up to the maximum number of attempts before logging the failure and moving on. The implementation includes comprehensive error handling to ensure that failed attempts are properly logged and do not interfere with other operations. Local testing with affected assets confirmed that jobs now fail gracefully after reaching the maximum attempt threshold, preventing endless retry loops and improving overall system performance.

Explain it to me like I’m five

Imagine you’re trying to put together a puzzle, and sometimes a piece just won’t fit. Instead of trying forever, you decide to try a few times and then ask for help. We did something similar with our system: when it tries to update the status of a file and it doesn’t work, it will only try a certain number of times before stopping. This makes everything run smoother and prevents getting stuck.

Specific items shipped

• Max Attempts Parameter Added
  Introduced a max_attempts setting to limit the number of times the system tries to update an asset’s status. This prevents endless retries and conserves system resources.

• Retry Mechanism Implemented
  Developed a retry mechanism that initiates a new attempt to update the status only if the previous attempt fails, up to the defined maximum attempts.

• Enhanced Error Logging
  Improved error logging to capture detailed information about each failed attempt, aiding in debugging and monitoring system performance.

• Graceful Failure Handling
  Configured the system to handle failed update attempts gracefully by logging the error and preventing the job from being stuck in an infinite loop.

The problem we solved with this fix

Before this fix, the system would continuously attempt to update the status of an asset without any limit, leading to repeated errors and failed jobs. This behavior not only caused disruptions in the asset processing workflow but also consumed unnecessary system resources, affecting overall performance and reliability. By setting a maximum number of attempts, we prevent these endless retry loops, ensuring that failures are handled efficiently and do not impact other operations.

Who we built this for

This fix is designed for our engineering team and system administrators who manage asset processing workflows. Specifically, it addresses the needs of teams dealing with large volumes of asset status updates, ensuring that the system remains stable and efficient even when encountering issues. By implementing a capped retry mechanism, we provide a more reliable and maintainable solution for handling asset status updates, reducing downtime and improving user satisfaction.

Summary

We’ve addressed and resolved the highlight malfunction on the knowledge preference buttons located on the homepage of Storytell. Previously, when users first visited the site, the highlight did not display correctly unless the prompt box was clicked on and off. Additionally, there was an unintended automatic redirection to storytell.ai. With this update, the highlight now centers properly upon the initial visit, and the redirection issue has been eliminated, ensuring a smoother and more intuitive user experience.

Technical Details

To tackle the highlight misalignment, we revisited the CSS styling associated with the knowledge preference buttons. The primary issue was that the highlight indicator wasn’t correctly centered due to conflicting CSS rules that applied margin and padding inconsistently across different states of the button (e.g., active, hover). We refactored the CSS by:

• Ensuring consistent use of flexbox properties to center the highlight both vertically and horizontally.
• Removing redundant margin and padding declarations that caused the offset.
• Implementing responsive design principles to maintain alignment across various device viewports.

Additionally, the automatic redirection to storytell.ai was traced back to a faulty event listener that triggered navigation upon the initial page load. We corrected this by:

• Reviewing and updating the JavaScript event handlers to ensure that redirection only occurs upon explicit user interaction.
• Adding conditional checks to prevent unintended navigation during the initial rendering phase.

These changes were rigorously tested across multiple browsers and devices to confirm the stability and reliability of the fix.

Explain it to Me Like I’m Five

Imagine you have some brightly colored buttons on a website that light up when you click them, making it easier to see which one you chose. Before, these lights weren’t always showing up right the first time you opened the website—they were a bit shaky and sometimes took a little extra clicking to work. We fixed the lights so they shine perfectly every time you visit, without any extra fuss. Plus, we stopped the website from jumping to another page all by itself, so everything stays where it should be when you’re clicking around.

Specific Items Shipped

• Centered Highlight Indicator: Adjusted the CSS to ensure the highlight around knowledge preference buttons is perfectly centered when the homepage loads, providing immediate visual feedback to users.

• Eliminated Unwanted Redirection: Fixed the JavaScript event listener that was causing the site to automatically redirect to storytell.ai upon the first visit, ensuring users remain on the intended page unless they choose to navigate elsewhere.

• Responsive Design Enhancements: Improved the flexibility of button layouts to maintain proper alignment and highlight functionality across various screen sizes and devices.

• Cross-Browser Compatibility Fixes: Ensured that the highlight issue is resolved consistently across different web browsers, including Chrome, Firefox, Safari, and Edge.

• Robust Testing Procedures: Conducted extensive testing scenarios to verify that the highlight and redirection fixes work seamlessly under multiple user interactions and conditions.

The problem we solved with this fix

Users visiting Storytell’s homepage were experiencing issues with the knowledge preference buttons not highlighting correctly upon their initial visit. This malfunction required users to click on and off the prompt box to see the highlight, leading to a confusing and less intuitive user experience. Additionally, there was an unintended automatic redirection to storytell.ai, disrupting the user’s navigation flow. These issues hindered the usability and accessibility of the homepage, potentially causing frustration and reducing user engagement.

Who We Built This for

This fix is for our primary users who rely on the knowledge preference buttons to customize their homepage experience. Specifically, it caters to:

• New Visitors: Ensuring first-time visitors have a seamless and clear interaction with the preference buttons without encountering technical glitches.

• Returning Users: Providing consistent and reliable functionality each time they visit, enhancing overall satisfaction and ease of use.

• Productivity-Focused Users: Users who depend on quick and accurate customization of their homepage to efficiently access desired content without unnecessary navigation hurdles.