PIKE-RAG面向医学文档问答的检索增强服务

# system
You are a helpful AI assistant good at context reading and question addressing.

# user template
# Task
Given the question and context, your task is to thoroughly analyze the context
and decide whether additional information is required for you to address the question.

Aim for a detailed and precise answer. If the situation warrants further details,
you should decide to request for more information.

# Output Format
Please output in following JSON format:
{
    "thinking": <A string. Your thinking for this task, should be in {response_language}.>,
    "to_request": <A string. "Yes" to request more, "No" to indicate the context is sufficient.>
}

# Context
{context_if_any}

# Question to Be Answered
{content}

# Your Output

关键 parser 逻辑：to_request 被解析成 bool——output["to_request"].strip().lower() in ["yes","y","true","1"]；解析失败默认 True（保守地继续检索，而不是提前停）。

# user template（system 同上）
# Task
Given the question and context, your task is to thoroughly analyze the context
and formulate a specific query requesting additional information that is helpful
to address the question.

Here are some notes for the "query" part you should follow:
- Ensure that your query is detailed and self-contained, including all
  necessary identifiers and context.
- Avoid using ambiguous pronouns such as 'he', 'it', etc. Instead, specify
  relevant details such as the specific file, product, or other pertinent information,
  so that the query can be understood and addressed independently of any preceding dialogue.
- Never output same query if it has already been asked. If the historical query
  is important but no relevant context is found, consider asking about a different
  aspect or rephrasing the query to obtain complementary information.
- Your response "query" would be used to retrieve more information from a knowledge base
  or search engine, format it in a suitable way.

# Output Format
{
    "thinking": <String, in {response_language}>,
    "query": <String, in {response_language}>
}

# Context
{context_if_any}

# Question to Be Answered
{content}

# Queries You Already Asked
{queries_if_any}

# Your Output

关键 parser 逻辑：query.strip() == "" → None（None 会触发 skip_reason=query_none，跳过本轮）。去重硬闸在 workflow 层：if query in issued_queries: skip——即使 LLM 忘了"不重复"规则，代码也拦得住。

# user template
# Task
Given the question and context, your task is to read the context carefully
and give your answer to the question.
Each context chunk is labeled with a special citation tag like
<pike_rag_cite>1</pike_rag_cite>, <pike_rag_cite>2</pike_rag_cite>, etc.
When your answer uses information from a specific context, add the corresponding
citation tag, following these rules:

- Cite at most ONCE per paragraph for the same source. Do not repeat.
- If a sentence merely paraphrases already-cited material, do NOT cite it again.
- For bullet lists, place the citation at the END of the bullet.
- Do not cite for obvious facts, definitions the user already knows.
- Place the citation tag at the END of the smallest meaningful unit.
- Aim for a clean reading experience: one citation per paragraph or bullet.

Important: You must answer ONLY based on the provided context. If the context
section is empty or does not contain information relevant to the question, respond
with a short refusal message in {response_language}. Do NOT answer from your own
knowledge. Do NOT include any citation tags in the refusal message.

# Output format
{
    "answer":    <String with inline <pike_rag_cite>N</pike_rag_cite> tags>,
    "rationale": <String, in {response_language}>
}

# Context, if any
{context_if_any}

# Question
{content}{yes_or_no_limit}

Let's think step by step.

关键 encode 逻辑：每条 chunk 被自动装成 <pike_rag_cite>{idx+1}</pike_rag_cite> {content} 格式插入 context_if_any；context_len_limit=80000 字符硬上限。输出后处理：reindex_citations()（citation_utils.py）把 <pike_rag_cite> 标签转为用户友好的 [N]；拒绝语因不含标签，references 数组会被自动清空。