mirror of https://github.com/HKUDS/CLI-Anything.git synced 2026-09-01 15:36:07 +08:00

Files

T

Devin AI b0e360262b refactor: remove deprecated answer, similar, and session endpoints

- Delete core/answer.py (Exa deprecating /answer)
- Delete core/session.py (not Exa-specific)
- Remove find_similar() from core/search.py (Exa deprecating /similar)
- Remove answer, similar, session CLI commands from exa_cli.py
- Fix ReplSkin constructor call (wrong kwargs)
- Add x-exa-integration header to all API calls
- Update all docs and tests to match
- Update setup.py and registry.json descriptions

Co-Authored-By: teo@exa.ai <teodorolgonzalez5@gmail.com>

2026-04-05 09:19:23 -07:00

2.8 KiB

Raw Permalink Blame History

EXA — Architecture & Design

Overview

cli-anything-exa is a CLI harness for the Exa API, an AI-native search engine built on neural embeddings rather than keyword matching. This harness makes core Exa capabilities available to any AI coding agent (Claude Code, Codex, OpenCode, etc.) via a structured command-line interface.

Backend

Exa exposes a REST API wrapped by the official exa-py Python SDK. Authentication uses EXA_API_KEY from the environment — no server process is required.

Command Hierarchy

cli-anything-exa [--json]
├── search   <query>          Neural/keyword/deep web search
├── contents <url> [url …]   Fetch full-text or highlighted page content
└── server   status           Verify API key and connectivity

Output Strategy

All commands emit human-readable output by default and structured JSON when --json is passed at the root level. JSON output is the recommended mode for agent pipelines.

JSON result shape for search/contents:

{
  "results": [
    {
      "title": "...",
      "url": "...",
      "published_date": "...",
      "author": "...",
      "highlights": ["..."],   // when --content highlights
      "text": "...",           // when --content text
      "summary": "..."         // when --content summary
    }
  ],
  "cost_dollars": {"total": 0.005}
}

Design Decisions

highlights as default content mode — Exa highlights are 10× more token-efficient than full text and are sufficient for most agent retrieval tasks. Full text is available via --content text when needed.

Research commands deferred — Exa's async deep researcher (/research/v1) has a start→poll→get lifecycle that warrants a separate v2 PR with proper state persistence.

Category slugs use hyphens — CLI uses research-paper, personal-site, financial-report (hyphenated) for shell-friendliness; the backend maps these to the API's space-separated values.

File Layout

exa/agent-harness/
├── setup.py
├── EXA.md                         (this file)
└── cli_anything/exa/
    ├── __init__.py
    ├── __main__.py
    ├── exa_cli.py                 Entry point, Click command tree, REPL
    ├── README.md                  Setup and usage guide
    ├── core/
    │   └── search.py              web_search(), get_contents()
    ├── utils/
    │   ├── exa_backend.py         SDK client init, contents/category helpers
    │   └── repl_skin.py           Shared REPL terminal UI
    ├── skills/
    │   └── SKILL.md               Agent-discoverable skill definition
    └── tests/
        ├── TEST.md
        ├── test_core.py           Unit tests (no API calls)
        └── test_full_e2e.py       E2E tests (real API, requires EXA_API_KEY)

2.8 KiB Raw Permalink Blame History Unescape Escape