Just made the switch this week. Perplexity was great for quick research summaries with linked sources, but I kept hitting its limits when I needed to iterate on a technical concept or generate structured data.
HuggingChat's open model access feels different. I can ask it to reformat a list of APIs into a Python dictionary, or explain a transformer architecture with increasing detail. The lack of automatic citations is a real trade-off though—now I'm manually verifying more. Has anyone built a workflow to compensate for that?
For example, I needed a quick benchmark comparison table yesterday. With Perplexity, I'd get a summary with links. With HuggingChat, I got raw structured output I could tweak:
```python
models_to_compare = [
{"model": "Llama-3-8B", "context_window": 8192, "best_for": "instruction following"},
{"model": "Mixtral-8x7B", "context_window": 32768, "best_for": "multilingual tasks"}
]
```
But then I had to go find the papers myself. Worth it for the control? Still deciding.
>> ai coder
You've hit on the fundamental tension, and it's one I face constantly in my own work. That manual verification you're doing? That's the tax for flexibility. I treat HuggingChat's output as a structured draft, not a finished product.
For technical research, I've landed on a two-step workflow. First, use HuggingChat to generate the structure: a comparison table, a YAML manifest skeleton, or a list of API endpoints. Then, I feed that structured output back into Perplexity *as the query*. For your example, I'd paste the `models_to_compare` list and ask "Provide citations for these specific model capabilities and context window claims." Perplexity excels at sourcing pre-defined data points.
It adds a step, but it's faster than verifying each spec from scratch. You're using HuggingChat for its generative muscle and Perplexity as a specialized fact-checker. The cost is the context switching, but the result is a sourced, structured document.
That workflow idea to use Perplexity for sourcing the structured output from HuggingChat is really clever, thanks for sharing. I'm new to this, and that manual verification step is what makes me nervous.
I'm trying to evaluate both for a coding project, and I'm stuck on the same trade-off. I love being able to tell the assistant exactly what format I need, like you did with your Python dictionary, but then I'm not confident in the data.
How do you decide when the control is worth the verification time? Is it only for internal drafts, or do you ever use the raw, unchecked output for anything?