Keep the queries short, no more than 3-5 words long. Focus on a single task and a single species, and include the species name in the query. This allows a more focused context to be presented to the LLM. Here are a few examples to get you started:
- What is Saigona baiseensis?
- Describe Saigona sinicola
- Describe Carvalhoma malcolmae
- Does Rhinolophus sinicus live in caves?
- Where does Laephotis botswanae roost?
- What cockroach species are there in India?
- What distinguishes Choanolaimus sparsiporus?
- What is the etymology of Gammanema lunatum?
- Are there any bats in Southern China?
- Describe Amnestus sinuosus
- What distinguishes Cynodon gibbus from other cofamilial genera?
About
Zai is the Zenodeo ai providing a Q&A access to the ~1M treatments extracted by Plazi and made available on TreatmentBank and BLR on Zenodo.
Zai responds to simple and focused questions with easy-to-understand summaries or pointed answers. It cannot (yet) respond to questions that span many documents, and is best at extracting answers from a single treatment that can be pin-pointed with a full-text search.
Zai is also the Chinese verb ε¨ (zΓ i) for describing existence in a location, similar to how we say in English, "to be at" or "to be in."
Check out more info on how this works.
Semantic Cache
We use the Supabase/gte-small model for generating vector embeddings of queries.
βββββββββββββ
βwhat is theβ
ββββββββββββββββββββββββββββββββΆβmeaning of ββββ
β βlife? β β
asks βββββββββββββ β
β β β
β β β β β β β β β β β β β β β βββ β β β βΌ
β semantic cache β ββ
ββββββ β βββββββββββ ββββββββΌβββββ β
β\O/ β βsearch β βconvert β ββ
β | β 42 β βfor exactβ βquery to β β
β/ \ βββββββββfoundββ€key in βββββunique β ββ
βuserβ β βcache β βcache key β β
ββββββ ββββββ¬βββββ βββββββββββββ ββ
β² β β β
β not found ββ
β β β β
β ββββββΌβββββ ββ
β β βconvert β β
β βto β ββ
β β βembeddingβββββββββββββββββββββ
β βvector β β
β β βββββββββββ
β β
β β ββββββββββββββββββββββ
β βsearch for query β β
β β βwith highest cosine β
βββββββββββfoundββsimilarity above a β β
β βpreset threshold β
ββββββββββββββββββββββ β
β β
β β β β β β β β β βΌ β β β β β β β β β β
no
β
βΌ
How this works
βββββββββββββ
βwhat is theβ
ββββββββββββββΆβmeaning of β
β βlife? β
asks βββββββββββββ
β β
β β β β β β βββ β β β β β β β β β β β β β β
β β semantic cache^ β
β β βββββββΌββββββ
ββββββ yes βis the β ββββββββ β
βuserβββ42βββΌβββββanswer in ββββββββββββΆβcache β
ββββββ βcache? β βββββ²βββ β
β βββββββββββββ β
no β β
β β β β β βββ β β β β β β β β β βββ β β β
β β β β β β β β β β βΌ β β β β β β
vector db ββββββββββΌββββββββ β β
β βconduct a vectorβ βββββββ΄βββββ
βsearch of the β β β response β
β βquestion to β βββββββ²βββββ
βidentify relatedβ β β
β βtreatments β β
ββββββββββ¬ββββββββ β βββββββ΄βββββ
β β β LLM β
βββββββββΌβββββββββ β βββββββ²βββββ
β βcreate a contextβ β
βfrom the β β β
β βfulltext of the β β
βββββββββ¬βββββββββ β β
β β β β β β β β β β βΌ β β β β β β
β β
βββββββββΌβββββββββ β
βquestion: "what β β
βis the meaning β β
βof life?" β β
β ββββββββββββββ
βcontext: top β
βranked paper β
ββββββββββββββββββ
A vector search using usearch allows searching for documents relevant to the question. The question is converted to a vector embedding and similar embeddings (using cosine similarity) are located in a db of ~3.5M embeddings generated from the 1M+ treatments. The fulltext of the topranked treatments is used to build a context for the question.
The LLM uses the context to generate the answer. The generated answers are stored in a semantic cache for fast retrieval for subsequent queries.
We are testing various LLMs including Alibaba's Qwen 3:0.6b, Meta's Llama 3.2:1b and 3b, and Google's Gemma 3 to generate the answers.