Zemi

Appearance

Models ​

Embeddings ​

qwen3-embedding:4b through Ollama, in five layers: paper (title and abstract), section, passage, section summary, paper summary. Search fuses the rankings from each layer and a keyword index.

Summarizers ​

Summaries, classification and judging are done by summarizers named in the config:

toml
summarizer = "deepseek-v4.1-flash"       # summaries and classification
code_summarizer = "deepseek-v4.1-flash"  # codebases
record_summarizer = ""                   # records; default: summarizer
judge = "opus-agent-pdf"                 # comparing summarizers
tutor = ""                               # the tutor; default: summarizer
tutor_effort = "medium"                  # how hard it thinks: low, medium, high
notebooker = ""                          # the agent making notebooks; default: summarizer
analyst = ""                             # writes about the paper of the day; default: summarizer

[watch]
judge = ""                               # rates new arXiv papers, picks the paper of the day; default: summarizer

Two kinds:

  • API: one request per section to an OpenAI-compatible endpoint (OpenCode Zen, OpenAI, Ollama's /v1).

  • Agent: an agent supporting ACP (Claude Code, opencode, Codex), run through Kaimon; reads the whole paper before summarizing. Needs Kaimon loaded in the same process (the zemi command, the Kaimon extension, Slate notebooks).

Built in:

  • deepseek-v4.1-flash, deepseek-flash: OpenCode Zen (opencode auth login).

  • gpt-5.4-mini: OpenAI, its key in ~/.config/zemi/openai_key.

  • opus-agent-pdf, opus-agent-text, opus-agent-code, sonnet-agent-pdf, sonnet-agent-code: Claude through a Claude subscription.

Add others:

toml
[[summarizers]]
name = "local-agent"
kind = "agent"
model = "ollama:qwen3:32b"      # or "sonnet", "acp:opencode:opencode/glm-5.3-flash", ...
reads = "text"                  # default "pdf", for models that read PDFs

[[summarizers]]
name = "local-api"
url = "http://localhost:11434/v1"
model = "qwen3:32b"
key = ""                        # or "file:~/.config/zemi/key", "env:API_KEY", "opencode"
request = { temperature = 0.2 }
price_in = 0                    # dollars per million tokens, for the cost estimates
price_out = 0

Changing summarizer marks existing summaries out of date; the next sync rewrites them.