BIG F.E.N.N.E.C. Update
F.E.N.N.E.C. is pretty cute and pretty helpful, but as of today he’s even harder, better stronger, faster. I have updated the knowledge base with a lot more content, and I also made Librarian Mode be even more in-depth by increasing his response length slightly. If all of that sounds like nerd speak to you, don’t worry about it :P The point is he’s even more dialed in and now supercharged with my entire library, as several of the latest content (this year) were not included in his training.
So what new information did F.E.N.N.E.C. devour? Let’s see:
All newer content in 2026 that has been released or will be released (through summer of 2026)
Pruned some content that wasn’t as useful or outdated
Corrected some key places in the transcripts where I had misspoken, and added various correction posts I’ve made over the years to the training data for more accuracy
Added another 150+ discussion questions for both discussion mode and for training (total now 650+ in Discussion Mode)
Added my landmark theology book, The Seven Transformations: Finding the Narrow Road by Pursuing a Sovereign Christ
Tweaked prompt to allow 3-6 rich paragraph responses (up from 3-5)
Re-chunked all my books and discussion questions (hundreds of thousands of words) from about 700 words per chunk to about 1000 words per chunk. Why this matters is that discussion answers and books are more dialed in and focused, so context is better communicated with larger text chunks for semantic indexing. When I talk on a podcast, I can meander through topics, so a 700 word chunk limit is fine. I decided to re-do this to make these resources even more relevant and powerful for FENNEC’s brain :P
Pruned 79 discussion questions that were short answers and redundant in the library (i.e. short answers to basic dispensationalism questions or situational responses to people’s lives that are addressed deeper in other places), and went through all questions and tweaked the shorter ones that I DID keep with a little clearer content for training purposes, like reiterating the question so FENNEC knows what was asked (only answers get logged into training) and specifying bible verses where necessary so those are indexed (i.e. someone would ask me about a particular verse, and the answer was something like “In this verse” — so I changed “In this verse” to the actual verse, so that it’s tagged in training). Also clarified some key answers so that they are more relevant. All of the questions are still in discussion mode, but for training the Librarian several were pruned and refined to improve answer quality.
Processed all (hundreds) of transcripts for transcription errors, misspellings, etc. for better accuracy and less chance of FENNEC getting confused. This took quite some time, but with the power of Claude I was able to resolve thousands of errors and make the training data even better.
Trained FENNEC on verse blocks from Berean QA and my various presentation notes, distilling everything into high density verse blocks by category so that you can ask him “give me the best Bible verses for _____” and he can now give you a list of verses because he can pull directly from these high-density chunks. Before, FENNEC relied on transcript data, which is filled with commentary and also reading of the verses most of the time (and sometimes transcription errors too unfortunately) which doesn’t give you the same density for a request like this. But now this information is in there. The topics are generally the 7 major areas of theology, as I outline in my recent book, The Seven Transformations. I have included a note now when you enter Librarian for the first time with a link to the FENNEC about page, where these categories will be listed in full. I hope this addition helps you in your conversations as well as personal bible studies :)
As a trade off for all this, and in order to keep FENNEC running as efficiently as possible, I had to make the decision to remove voice output. I know! “Friendly Sidekick” was totally FENNEC’s voice. But just to give you an idea, using voice mode consumes exponentially more than text output. In just about 2.5 months, voice mode alone accounts for something like $450 of usage between a dozen or two conversations. There is no way for me to charge a premium or gate this feature to those who want to use it. I could either go with some crappy service like Amazon, but his voice would be dull and robotic, or I could cut the feature entirely. So I have decided to do the latter. Claude 3 has also been deprecated by Anthropic, so all the models were upgraded (Claude 3.5 —> Claude 4.5), so FENNEC’s abilities are better but they also cost 25% more because of that. It’s not a huge problem for text only, but when you add voice to the mix it’s just not sustainable sadly. So we will have to wait until the technology evolves and either becomes cheaper or it can be some kind of premium perk somehow that can be gated for those who really want to use it. In the meantime you can still voice to text (your voice into text, as in speaking your question and it transcribes), so that’s still available.
A few other important points:
A statement of faith about what I believe and how I label my positions on the 7 major doctrinal areas (soteriology, christology, thanatology, ecclesiology, eschatology, cosmology, the Sabbath) has been added to FENNEC’s prompt. This wasn’t so much important for answer questions as it was as a guide against interpreting KB (knowledge base) results and labeling me with mainstream labels. The AI may read, for example, that I believe the prophecy about Damascus was fulfilled in the past while answering your question about Zechariah 14. It may then assume that I am “preterist-leaning” and label my position, even cautiously, but this is incorrect. That’s not a KB problem, that’s a framing problem, and that going forward has been fixed so that FENNEC understands my positions on every question and frames my views correctly without appealing to existing denominations or similarities with others (i.e. “Calvinist-like” or similar comparisons).
FENNEC’s loading graphic (as the answer is being prepared) now has a “sniffing for answers” bar that will be displayed rather than a generic loading wheel
The automated reminder notification that goes off during discussion mode or librarian mode (to remind you that you can also do quizzes) is now a little more clear that it is a friendly reminder that you have other modes of learning available — it is NOT an indication the chat has been reset or that you need to hit “library” again.
On follow-ups: When tweaking some things on the back end, I removed a feature that Voiceflow has by default which is to import the memory of the conversation into the chat. This is actually not needed given FENNEC’s architecture because every question gets treated as a new situation, or a follow-up of the previous offer he gave you (“Would you like to explore XYZ further?”) FENNEC will always give you a follow-up question, and the system extracts that into a meaningful query so that if you affirm (“yes, sure, go ahead” etc.), the system takes that query and pulls library results. However without going into complicated mechanics, this mechanism of saying “yes” to a follow-up was broken when I removed the memory feature. I did this because it was eating into the total credits unnecessarily. In other words, all those turns being saved to memory were being sent to the LLM every time, which is a massive waste of tokens. But removing this feature unmasked several bugs. Long story short, these should be fixed. You should be able to change topics normally by asking a normal question, and then if you affirm, he will take his follow-up question and look it up for you. If you are too verbose or for whatever reason you get a “no match” response (i.e. “Sorry, I didn’t get that” etc.) then there is a note there now to tell you to say “yes” if you meant to simply agree to his previous follow-up. Doing so will jump back into the conversation fluidly.
On Hallucinations and Interactions: The more detailed your questions are, the better FENNEC can help you. One word questions or obscure questions are not going to be useful because they can lead to inconsistent answers or even hallucinations. This is rare, but a good rule of thumb is to always put in more if you want more. If there’s no match in the KB, FENNEC is supposed to prompt you to ask a different question or a better question. Occasionally, rarely, he may hallucinate on something. “Hallucination” means he will invent something based on standard internet knowledge (which of course is the mainstream evangelical position). This has been very rare, and most of the time even when it does happen — it is a minor issue interwoven with a 95% correct and edifying answer. Always clarify a point of confusion, and if it’s a serious problem please document your conversation with the following and email it to me: date and time you were talking, the exact question you asked, the answer you got (screenshots) and what the issue was, as well as how long you were in that particular session (5 minutes, 2 days? etc.) Most of the time why this will happen is because of either a transcription problem or because of how I said something. In other words, I may have said something in a confusing way — something that if you were listening to me and had context, you would understand what I meant. But this gets lost in text, especially with the transcription layer, so in certain cases these “polluted” strands can confuse FENNEC and make him say something that I don’t actually believe. It may also be the case that certain chunks caught me reading from an article or perspective that does not align with my views (because I was documenting it), and this needs to be pruned out because it pollutes the data. These issues are rare but they happen, and working out the kinks can be a real “needle in the haystack” process, which is why I need your help to pay attention and document them over time. I have also tweaked his prompt to add extra insurance, in the sense that he should never assume I hold the mainstream evangelical view, and there’s a note now when you enter librarian to contact me if there is an issue. Either way, despite these things I’d say the overall response quality is pretty high, and FENNEC is an amazing tool to help you use my library. But he’s not 100% perfect as nothing is, so just be mindful.
Overall the quality of responses and references should be EVEN BETTER than before, and discussion mode is now complete. I do not foresee adding to these training materials any time soon, but we will see how things go and if there are key things that FENNEC can’t address then I will fill them in as time goes on. There is now enough content to last you years of research and study, so make it an adventure! I checked the stats and hardly anyone is using Quiz or Discussion mode, which is sad because these are great and fun tools to test your knowledge. So if you haven’t yet checked out F.E.N.N.E.C., then don’t be shy and give him a try — he doesn’t bite too hard.