I tested with the system prompt you shared, and you are right about the gemini-exp-1206
model, sometimes it misses to provide grounded sources. I will escalate the issue to the team.
Have you tried with gemini-1.5-pro
or gemini-2.0-flash-exp
? They seem to provide grounded responses when I tested them.
3 Likes
Thank you for the confirmation and for reporting the issue to the development team. I have evaluated the gemini-1.5-pro
and gemini-2.0-flash-exp
models as suggested. My preliminary testing indicates that while they appear to handle response grounding more effectively, their overall performance, in terms of semantic accuracy and contextual coherence, currently lags behind gemini-exp-1206
by 8.7%, based on a benchmark of 1500 queries representative of my operational needs. Furthermore, gemini-2.0-flash-exp
, despite being 12.3% faster, currently exhibits limitations in context length, rendering it unsuitable for approximately 25% of my typical requests. Regarding Think 1-21
, it unfortunately lacks Google Search grounding functionality, a critical requirement for ensuring information reliability and verifiability, indispensable to my workflow. I will continue to monitor the development of these models and conduct thorough testing to assess any future improvements that might render them suitable for my operational needs. However, for now, gemini-exp-1206
, despite its current limitations in sourcing, remains the optimal choice to ensure the level of accuracy and reliability demanded by my application.
1 Like
I appreciate your dedication to resolving this issue. Please promptly notify me upon implementation and validation of the update addressing the grounding of sources for the gemini-exp-1206
model.
1 Like
I have also noticed that the System Istructions are less reliable than sending the System prompt in the First message of the chat.
@GUNAND_MAYANGLAMBAM Hi Gunand,
I wanted to update you on the issue related to the Google Search grounding functionality. The previous issues primarily concerned the malfunctioning grounding in Gemini Experimental 1206
. The key problem, and this is crucial, is that the exact same issue also occurs with Gemini 2.0 Pro Experimental 02-05
. I’ve run similar tests, and grounding works 90% of the times, only in the first response message; subsequent responses do not use Google Search for updating information, even when explicitly requested end enabled in the slider.
Given that the problem reappears in Gemini 2.0 Pro Experimental 02-05
, it’s clear that this isn’t an isolated issue with the 1206 version. I kindly request that you extend the escalation of the problem to the team, highlighting the persistence and the limitation to only the first response message.
Thank you for your attention and assistance.
I would add also that the quotes happens to cute Words and phrases in the middle instead of at the end of the sentence.