Perplexity
I am using Perplexity Enterprise. Has anyone any experience with this AI application and Devonthink
No.
Have you? Please share.
1 Like
In my experience it’s only useful for doing a deep research or as a search backend (see Settings > AI > Search). For all other tasks and to work with your documents & databases other providers are highly recommended instead. The new MCP tools and skills of v4.4 are also not supported.
It is possible to connect Perplexity to the Devonthink MCP connnector if you use Warplet or other methods to convert Devonthink’s local MCP server to HTTP.
Overall Perplexity works really well if you want it to autonomously perform very deep research on a particular question.
Claude Cowork / Claude Dispatch on the other hand are much better if you want to set up a complex and/or repeatable workflow involving multiple steps or specific formatting or combining AI with classic coding languages.
Define very deep. I usually prefer the results of the Claude and ChatGPT apps. More substance, less nonsense and bloat. But YMMV.
Perplexity can prepare exhaustive literature surveys on an academic topic yielding a 20-30 page report with hundreds of references - the level of depth one might expect in a graduate level overview of a topic. [I am not suggesting that is usable for publication or without confirmation - but if you are working on a given topic and want some initial framing or you want an additional literature review source to help assure you have included everything pertinent, Perpelxity is an amazing tool.]
On the other hand, if you are looking for practical advice on how to take better photos with your iPhone then Perplexity deep research is going way overboard.
Most likely most frontier models supporting search could do this too. The difference is the prompt, e.g. Perplexity’s system prompt is of course optimized for searching. Perplexity doesn’t use proprietary models, e.g. in the past they definitely used Claude. Not sure what they’re currently using.
For truly deep research I believe peformance is a function both of the agent harness and the model(s) that the harness calls. Deep research typically involves multiple subagents making LLM calls in parallel; the agent harness facilitates this.
The agent harness in particular used by Claude Dispatch and Claude Cowork is optimized stunningly well for workflows involving document processing and/or knowledge research. It identifies edge cases autonomously and proposes solution options to the user. It finds errors autonomously and prompts the user with options to correct the errors. It suggests third-party databases or APIs that may be useful additional sources of information. While it makes errors and thus its work always needs to be confirmed, it functions at a level comparable to a graduate student engaging in research to assist a professor.
Perplexity’s agent harness works best if you want depth from a known source of information. Claude’s agent harness works best if you want analysis to identify edge cases or errors or if you want help on where to look for unusual facts.
For researching I actually prefer Claude Chat, not Cowork. Same backend but optimized for different usage scenarios.
It all depends on the deliverable you are seeking and how complex the research is.
Cowork is better if you want it to read/write local files. Also Cowork can run parallel sub-agents whereas Chat runs them sequentially. Cowork can also use a browser to access sites behind a login.
Chat works well as a quick/easy session in the chat window.
i have used perplexity with DEVONthink using a applescript which allows it to search through a specific database. I does a very good job. I use it in connection with a litigation matter and it does a better job than most paralegals or junior lawyers can do. When it searches the devonthnk database it does not use the space repository created for my case. So i might be a problem if it cant access both the space and database at the same time.
1 Like
i have now connected DEVONthink to perplexity using MCP Warplet . It was difficult but managed to resolve issues after a few hours . It seems that it will be much more useful than the applescript method. All in all, I found Perplexity and using Fable totally mindboggling
1 Like
And very deep for me means dumping legal proceedings, evidence, unfiled documents , emails, transcripts and having AI draft a 20 page submission to the court analyzing contractual provisions, seeking out evidence to support the position and fromating the argument with footnotes at the bottome of the page, not endnotes.
The legal fees could be $25,000 to $50,000 for a senior lawyer to do it with Fable, while without AI the client would pay $150,000 and up. Fable cost about 7,000 $ and lawyers time $75,000. So you would assume legal costs should come down. I doubt they will but the efficiencies here are obvious . I suspect law firms wiill find ways to keep those bills high.
The funny thing is when I asked Perplexity if a sole practictioner using Fable was at a disadvantage with a large firm uisng Harvey and the short answer is no. A small firm now can leverage AI to compete with largefirms. That is my belief as well as a result of the work I am doing uisng AI Perplexity and the Anthropic Fable model.
Or the costs will remain the same but the quality will go up.
AI is an amazing tool - but in law or medicine or any other profession you still need someone with extensive domain-specific experience to take responsibility.
As great as AI is, even the best frontier models can hallucinate. Worse, they hallucinate in a highly convincing way.
One AI fabrication - if submitted to a court as authentic - can blow a whole case.
So the way the economics of law firms work, I believe the costs will remain the same, the more senior lawyers will find a way to charge alot and work less. The quality will improve or at least hight quality lawyers will continue to produce high quality work.
I dont understand how hallucination can happen as the senior lawyer will , assuming he is a high quality lawyer to begin with, will carefully review the work of AI as it evolves. There should not be any AI fabrication heading to court, unless the lawyer is of lower quality or lazy. The role is not to let AI run the case but to get it to act as a brilliant 3 year experienced lawyer assisting the. preparation . It drafts memos in response to questions I put, redrafts arguments in clear language with no typos , with charts etc… better than my paralegal ever did.
1 Like
The fact that I, working far from any legal sphere, know that lawyers have already been:
- Submitting briefs with hallucinations within them
- Been caught doing so
- Been publicly exposed as doing so
Would lead me to conclude that no one should be expecting the positive aspects of human behaviour to win out in this situation.
3 Likes
“Should not” and “will not” are not the same. Certainly there are abundant examples of lawyers embarrassing themselves by letting AI hallucinations appear in court documents.
That’s the trap of AI generally. 90% accuracy is good enough that mistakes will be rare. Potentially rare enough for someone under deadline pressure to say “well, I don’t have time for this right now.” And probably it even works, so there’s even less incentive to check everything the next time. But eventually even a 10% chance comes up.
There’s also, at this point, abundant evidence of AI-related deterioration of skills. The less you engage with the material, the less capable of engaging with the material you become.
There are good lawyers and not so good lawyers - that has been that way for eons. It is true in every profession and every other facet of life.