first draft

[donaldziff]

There is a long way to go on this!

[hwchase17]

this is awesome @donaldziff !

how are you thinking about merging this in? im pretty fine to merge in and publicize and get others feedback... unless you wanted to try to formalize/generalize some parts?

[donaldziff]

If this were to go in soon, I would like to make some minor tweaks. But the current code is simplistic, and I don't want to represent it as finished, so I'm a bit embarrassed to put it in as it is.

In the longer term I'd like to:

• add a database so the AI can find properties by embedding search rather than full-text search
• add a semantic element to item search: full-text search works better for q-items than for properties, but it can still get confused
• add a query planning stage at the beginning. The AI needs to think about the graph structure around the specific objects before generating sparql

But I would love to get feedback, and if getting this in sooner would help, then please give me a chance to make tweaks, and let's get it in! Here are some short-term questions:

• If it were merged soon, would the place I chose (docs/modules/agents/agents/wikibase_agent.ipynb) be the right one?
• I would like to formalize/generalize and write proper code like langchain/agents/agent_toolkits/sql.. But if I could get feedback sooner, that would be great. Should I enumerate some of my open questions in the notebook?
• What forum do you recommend for feedback? A github "discussion"? Discord? Should I invite that in the notebook?
• I am quite puzzled why the agent performs fairly predictably using davinci but becomes unpredictable using turbo or 4. It seems that those models hallucinate tool steps that are not among the possibilities I list. Do you have any ideas about that? Can this be solved by prompt engineering?

Thanks for your attention @hwchase17 !

[hwchase17]

• If it were merged soon, would the place I chose (docs/modules/agents/agents/wikibase_agent.ipynb) be the right one?

Yeah that place seems good for now!

• I would like to formalize/generalize and write proper code like langchain/agents/agent_toolkits/sql.. But if I could get feedback sooner, that would be great. Should I enumerate some of my open questions in the notebook?

Yeah questions in notebook sounds great. My main questions are what parts are generalizable vs specific to this data?

• What forum do you recommend for feedback? A github "discussion"? Discord? Should I invite that in the notebook?

i dont think we have a great/standardized way, so for now I would say whatever is most convinient for you

• I am quite puzzled why the agent performs fairly predictably using davinci but becomes unpredictable using turbo or 4. It seems that those models hallucinate tool steps that are not among the possibilities I list. Do you have any ideas about that? Can this be solved by prompt engineering?

probably prompt engineering... have heard this from other ppl as well

[hwchase17]

should we merge this in? updated thoguhts?

[donaldziff]

Yes, it's ready to go! Thanks for your patience

[hwchase17]

Yes, it's ready to go! Thanks for your patience

thank you! will share it tmrw