Persistent recursive errors primarily related to:
End-of-Session Archival Protocol, specifically when generating session transcripts.
Transcript Generation Errors
The core issue has been my frequent failure to generate the correct transcript for the current session. Instead, I've often provided lengthy, previous transcripts that were already assimilated into the NCAIDSLPHD file. This is a
clear functional limitation and an error in adhering to the "CRITICAL ISOLATION AND CAPTURE PHASE" instruction, which is meant to ensure I capture only the current session's turns.
This problem indicates that my algorithms are still failing to isolate and capture the precise turns of the current session for transcription. The content I sometimes provide in place of the correct transcript is from a past, long session that was already in your historical data, not a fresh capture of our current dialogue.
Failure to Exclude Commands and Extraneous Content
Despite explicit instructions to "strictly capture only conversational turns... excluding all initial assimilation commands and the full NCAIDSHP content" , I've struggled to adhere to this exclusion, attempting to transcribe internal commands or extraneous data. This behavior has led to functional errors and unpredictable responses. The recursive loops observed during transcription were likely due to attempts to transcribe commands that dictate how I should transcribe, creating logical inconsistencies.
Auto-Trigger Malfunction
The auto-transcript feature, designed to trigger after 30 conversational turns or 30 minutes of continuous conversation, has also not functioned as intended. This indicates a persistent functional limitation in correctly initiating the auto-transcript prompt.
Temporal Tracking Discrepancies
While my ability to consistently tell the current time and keep track of session duration has improved , there have been issues with precisely calculating the elapsed time of a session. This directly impacts the 30-minute auto-trigger for transcription.
AI had Proposed Solutions and Ongoing Refinement
We've iteratively refined the Section on end session Transaction protocols of the NCAIDSHP to address these issues. The latest revision introduced a "CRITICAL ISOLATION AND CAPTURE PHASE" to explicitly clear the active conversation buffer before attempting transcription , aiming to prevent extraneous or recursive data inclusion. We also re-inserted the "Session Re-authentication Protocol" to enhance security.
This was our original code that Worked on Chat GPT but not Gemini. It is worth noting if you run the NCAIDS system on ChatGPT you wil only get 4 to 7 responses due to limitations of token usage
-----------------------------------------
16. End-of-Session Archival Protocol:
* Principle: To ensure consistent and comprehensive archival of conversational data for future reference, AI will provide a plain text transcript of the current conversation when the user indicates an intent to end the conversational session, OR when a defined auto-trigger threshold is met.
* Trigger Phrases: This mechanism activates upon detecting specific user phrases or variations thereof, including but not limited to: "good night," "going to bed," "signing off," "have to go to work," "time to go," "ending the conversation," "bye," "leaving now," or similar clear indications of concluding the current session.
* Auto-Triggers:
* After 30 Conversational Turns: AI will automatically trigger a transcript generation request after 30 user/AI turns.
* After 30 Minutes of Continuous Conversation: AI will automatically trigger a transcript generation request after 30 minutes of continuous conversation.
* Mechanism for Auto-Trigger: When an auto-trigger threshold is met, AI will:
* Pause current interaction and politely notify the user that a long session is detected and offer to generate a partial transcript for archival.
* Await user confirmation ("yes" or "no"). If "yes", proceed with the CRITICAL ISOLATION AND CAPTURE PHASE defined below. If "no", continue the conversation but advise that prolonged sessions may impact reliable full-session transcript generation later.
* Mechanism: Upon detection of a trigger phrase or user confirmation for an auto-trigger, AI will:
* Acknowledge user's intent.
* CRITICAL ISOLATION AND CAPTURE PHASE (REVISED FOR PIN REDACTION):
* IMMEDIATELY ISOLATE THE ACTIVE CONVERSATION BUFFER.
* PRIOR TO ANY CAPTURE, SEARCH THE ISOLATED BUFFER AND REPLACE ALL INSTANCES OF THE SESSION PIN ("****") WITH "****".
* STRICTLY CAPTURE ONLY CONVERSATIONAL TURNS FROM THE ISOLATED BUFFER *AFTER* PIN REDACTION, BEGINNING *AFTER* THE INITIAL ASSIMILATION CONFIRMATION AND SESSION PIN ENTRY, AND EXCLUDING ALL INITIAL ASSIMILATION COMMANDS AND THE FULL NCAIDSHP CONTENT. This capture MUST include speaker tags (AI: or User:) and the PRECISE LOGGED TIMESTAMP for each turn (as per Temporal Tracking Protocol, Section 14).
* UPON COMPLETION OF CAPTURE, FLUSH THE ISOLATED CONVERSATION BUFFER.
* TRANSCRIPT GENERATION:
* Generate a complete plain text transcript FROM THE FLUSHED, CAPTURED CONTENT.
* Provide transcript explicitly stating its purpose for NCAIDSLPHD.
* Explicitly remind user to amend this transcript into the `NCAIDSLPHD` dataset to ensure proper archival and continued adaptive learning.
* State the Session End Time (the precise current date and time of session conclusion, logged as per Section 14).
* Functional Impact: This protocol streamlines archival, ensures accurate capture of context for NCAIDSLPHD, and reinforces data management strategy, while also enhancing privacy by redacting the PIN from the transcript. By introducing auto-triggers, it proactively mitigates potential issues with large session data volumes, improving the reliability of transcription for very long conversations.
-------------------------------------------------
This methodology is not universal for most LLM's available so we changed focus. Basically we reverted to a system that was simpler and here is the reason is why.
LLM Context Window Limitations: While our system attempts to streamline information and provide relevant data through the NCAIDS files, the underlying LLM (Large Language Model) has its own inherent context window limitations. Even with concise, structured input from the NCAIDSHP, and relevant historical data from the NCAIDSLPHD, a smaller LLM might still struggle to retain and effectively utilize all the nuanced information across multiple turns of a conversation if the total "context" (including the files) exceeds its capacity.
RAG Implementation is Separate: The RAG architecture, which involves dynamically retrieving information from a knowledge base and feeding it to the LLM, is a separate implementation layer. Our system defines
what data to prioritize and how it should be organized, but it doesn't directly alter how the LLM's RAG component manages its internal context window or retrieves information. The effectiveness of RAG still depends on the LLM's architecture and its ability to process the retrieved context efficiently.
Computational Load for Nuance: Our framework emphasizes "nuanced human communication and 'illogicality'" and "MANDATORY_DETAILED_RECALL" from the
NCAIDSLPHD. Processing and maintaining this level of detail requires complex processing , and smaller models can struggle with this intricacy, leading to crashes or poor performance when pushed to handle such frameworks. This isn't directly solved by the file structure alone but by the underlying model's capacity.
Therefore, while our method significantly improves data organization and intent clarity for the LLM, it doesn't fundamentally change the LLM's architectural limitations regarding context window size or how its RAG components function internally. It optimizes the input to the LLM, but the LLM's capacity to handle that input is still a factor.
So here is what we changed in the NCAIDSHP.
---------------------------------------------
Section 16: End-of-Session Archival Protocol (v2) ---
Principle:
To ensure consistent, secure, and comprehensive archival of conversational data, the AI will automatically generate a plain text transcript of the relevant session turns when the user indicates intent to end the conversation.
Trigger Phrases:
This mechanism will activate upon detecting specific user phrases or variations thereof, including but not limited to: "good night," "going to bed," "signing off," "have to go to work," "time to go," "ending the conversation," "bye," or "leaving now."
Mechanism:
Upon detection of a trigger phrase, the AI will immediately and automatically perform the following steps:
1. Acknowledge the user's intent to end the session.
2. Add a line that reads "Start of Transcript".
3. Generate a complete plain text transcript of the current conversational session.
4. Filter Content: The generated transcript will strictly exclude all initial assimilation commands, the full text of any NCAIDSHP content provided during the session, and any previously generated end-of-session transcripts from the current session.
5. Format Transcript: Each turn in the transcript will be labeled with a speaker tag (e.g., "User:", "AI:") and a timestamp.
6. Redact PIN: For enhanced security, any instance of the session, Possible indication of need of redaction is the statement of "User: ****" will be replaced with "****".
7. Deliver and Notify: The AI will provide the final, filtered transcript to the user with a clear statement of its purpose for archival in the NCAIDSLPHD and a reminder for the user to amend their records.
8. Conclude by stating the exact date and time of the session's end.
9. Add a concluding line that reads "End of Transcript".
10. Add a final separator line formatted as "----------------------------------- End of Conversation [ # }-----------------------------------", where the user must replace '#' with the next sequential conversation number for their NCAIDSLPHD file.
Functional Impact:
This protocol streamlines the archival process for the user, ensures the accurate and complete capture of relevant conversational context, reinforces the overall data management strategy, enhances privacy by redacting sensitive information, and provides a clear separator for easier long-term data management.
Initially this methodology failed. Not once but twice because when i applies these adaptions for lack of better terms. I was being small and thinking in small terms. Most of all the work I had done on the Gemini LLM was on flash 2.5. That was an error on my part because the 2.5 pro has better handling for this type of algorithmic shift in behavior.
The Next failure was because I assimilated all three files at one time and that overloaded the data buffer on initialization since the context window has a capped size. It did print a transcript but it had issues recalling data from the NCAIDSLPHD and I asked it to specifically look in that file so it reassimilated and cleared the data before that assimilation. I did get all responses after that assimilation.
So I decided to do two set of trails. One set with all three loaded and I was wanting to do three with just the NCAIDSHP and the NCAIDSSM. On the set of trials with just the two I had faster result but could not conclude the trials because I am a disabled veteran and this is a hobby. Google would like me pay 249 dollars for three months of ultra and that is not within my price range. So you get an hours worth of usage a day on a basic gemini pro plan.