searchis-query

Searchis evidence retrieval

A semantic search tool that returns original-text evidence from internal financial research archives based on a natural-language question.

Pre-checks

npx @openduo/searchis status --json

Based on the activated field, decide the next step.

Not activated → guide the user

Output the following information andstop：

⚠️ Searchis 内部文档库未激活。

Searchis provides a semantic search service for internal financial research notes, earnings call transcripts, and industry analyses.

Activation steps:

1. Contact the administrator to get an invitation code
2. Run: npx @openduo/searchis activate <invitation code>

## 调用

```bash
npx @openduo/searchis query "<研究问题描述>"

The input is a natural-language question, which can be short or a long description with context. Searchis's internal agent will understand the question, break it down into an internal search strategy, retrieve relevant documents, and extract original-text excerpts.

A call usually takes 60-180 seconds to complete (the agent performs multiple rounds of internal search). The CLI uses a streaming protocol to continuously receive results, with continuous progress output on stderr indicating the agent is working. It only breaks off if the network or backend is truly stuck (default dial 30s, idle 60s); do not add extra timeout wrappers.

When programmatic parsing is needed, add --json to switch to JSON output (default is markdown with citations).

Output (default markdown)

The main content is a research answer with inline citations such as [1] [2] , followed by a Sources section listing each piece of evidence's title, author, date, line number, and original text. Citation numbers are indices within the file; when presenting to users, use readable forms such as [内部调研 {title} {date}] .

--json output

{
  "answer": "## 标题\n... [1] ... [2] ...",
  "evidences": [
    {
      "id": "abc123",
      "date": "2026-03-27",
      "title": "光芯片专家交流-2026年EML、CW光源全球供给情况...",
      "source": { "author": "Shirley" },
      "quote": "原文引用，未经改写",
      "lineRange": [40, 45],
      "context": "该证据的一句话描述"
    }
  ],
  "stats": { "duration_ms": 65000, "files_searched": 4 }
}

Field descriptions

• quote — text copied exactly from the original document, not rewritten by an LLM
• date — document date extracted from the file name
• title — title of the research note (from frontmatter)
• source.author — note author
• id — file hash, shared by multiple pieces of evidence from the same file
• context — a one-sentence summary of that piece of evidence
• answer — the service's internal draft organization of the evidence,for reference clues only: it may miss key evidence and may misread; the final conclusion is to be assembled by the caller based on evidences
• evidences: [] — no relevant evidence found (a valid result, do not fabricate)

Error handling

Return	handling
{"error":"not_activated",...}	show activation guidance
{"error":"query_failed","message":"...401..."}	prompt npx @openduo/searchis refresh
{"error":"query_failed","message":"...429..."}	wait 30s and retry
dial / idle timeout (network or backend stuck)	report timeout, suggest retrying with a more precise query

When using evidence

• Cite the original text from the quote field, do not rewrite
• Label the source [内部调研 {title} {date}]
• When results are empty, report truthfully, do not fabricate