lc_rag/graph_rag/
extractor.rs

1// src/retrieval/graph_rag/extractor.rs
2//! LLM-based entity and relation extraction from document text.
3//!
4//! Sends a structured prompt to the LLM, parses the JSON response into
5//! [`ExtractedEntity`] and [`ExtractedRelation`] lists.
6
7use lc_core::language_models::BaseChatModel;
8use lc_core::tools::ToolDefinition;
9use lc_schema::Message;
10use serde::Deserialize;
11use serde_json::json;
12
13use crate::structured::{chat_structured, StructuredChatResult};
14
15/// An entity extracted from text by the LLM.
16#[derive(Debug, Clone, Deserialize)]
17pub struct ExtractedEntity {
18    /// Entity name.
19    pub name: String,
20    /// Entity type (e.g. Person, Organization, Technology).
21    #[serde(rename = "type")]
22    pub entity_type: String,
23    /// Optional description of the entity.
24    #[serde(default)]
25    pub description: String,
26}
27
28/// A relation extracted from text by the LLM.
29#[derive(Debug, Clone, Deserialize)]
30pub struct ExtractedRelation {
31    /// Source entity name.
32    pub source: String,
33    /// Target entity name.
34    pub target: String,
35    /// Relation type (e.g. works_at, uses, part_of).
36    #[serde(rename = "type")]
37    pub relation_type: String,
38    /// Optional description of the relation.
39    #[serde(default)]
40    pub description: String,
41}
42
43/// The full LLM extraction response.
44#[derive(Debug, Deserialize)]
45pub struct ExtractionResult {
46    /// Extracted entities.
47    #[serde(default)]
48    pub entities: Vec<ExtractedEntity>,
49    /// Extracted relations.
50    #[serde(default)]
51    pub relations: Vec<ExtractedRelation>,
52}
53
54const EXTRACTION_PROMPT: &str = r#"You are a knowledge graph extraction assistant. Given the following text, extract entities and their relations.
55
56Return a JSON object with exactly two keys:
57- "entities": an array of objects, each with keys "name", "type", "description"
58- "relations": an array of objects, each with keys "source", "target", "type", "description"
59
60Rules:
61- "source" and "target" in relations must match entity "name" values exactly.
62- Keep entity types simple: Person, Organization, Location, Technology, Concept, Event, etc.
63- Keep relation types simple: works_at, located_in, uses, created, part_of, related_to, etc.
64- Extract at most {max_entities} entities and {max_relations} relations.
65- Return ONLY the JSON object, no other text.
66
67Example:
68Text: "Alice works at Google as a software engineer. She uses Python and TensorFlow."
69Output:
70{
71  "entities": [
72    {"name": "Alice", "type": "Person", "description": "A software engineer at Google"},
73    {"name": "Google", "type": "Organization", "description": "A technology company"},
74    {"name": "Python", "type": "Technology", "description": "A programming language"},
75    {"name": "TensorFlow", "type": "Technology", "description": "A machine learning framework"}
76  ],
77  "relations": [
78    {"source": "Alice", "target": "Google", "type": "works_at", "description": "Alice is employed at Google"},
79    {"source": "Alice", "target": "Python", "type": "uses", "description": "Alice uses Python"},
80    {"source": "Alice", "target": "TensorFlow", "type": "uses", "description": "Alice uses TensorFlow"}
81  ]
82}
83
84Text:
85{text}"#;
86
87/// 实体/关系提取工具定义(P2-1):强制 LLM 输出结构化 JSON 参数。
88fn extraction_tool() -> ToolDefinition {
89    ToolDefinition::new(
90        "extract_entities_relations",
91        "从文本中提取知识图谱实体与关系,返回 entities 与 relations 两个数组",
92    )
93    .with_parameters(json!({
94        "type": "object",
95        "properties": {
96            "entities": {
97                "type": "array",
98                "items": {
99                    "type": "object",
100                    "properties": {
101                        "name": { "type": "string" },
102                        "type": { "type": "string" },
103                        "description": { "type": "string" }
104                    },
105                    "required": ["name", "type", "description"]
106                }
107            },
108            "relations": {
109                "type": "array",
110                "items": {
111                    "type": "object",
112                    "properties": {
113                        "source": { "type": "string" },
114                        "target": { "type": "string" },
115                        "type": { "type": "string" },
116                        "description": { "type": "string" }
117                    },
118                    "required": ["source", "target", "type", "description"]
119                }
120            }
121        },
122        "required": ["entities", "relations"]
123    }))
124}
125
126/// 解析一次结构化调用结果:优先 tool_calls 参数,其次文本 JSON。
127fn parse_structured(result: &StructuredChatResult) -> Option<ExtractionResult> {
128    if let Some(args) = &result.tool_args {
129        if let Ok(parsed) = serde_json::from_value::<ExtractionResult>(args.clone()) {
130            return Some(parsed);
131        }
132    }
133    parse_extraction(&result.content).ok()
134}
135
136/// Extracts entities and relations from text using the LLM.
137pub async fn extract<M: BaseChatModel>(
138    llm: &M,
139    text: &str,
140    max_entities: usize,
141    max_relations: usize,
142) -> Result<ExtractionResult, super::GraphRAGError> {
143    let prompt = {
144        use lc_prompts::PromptTemplate;
145        let template = PromptTemplate::new(EXTRACTION_PROMPT);
146        let max_entities_str = max_entities.to_string();
147        let max_relations_str = max_relations.to_string();
148        let mut vars: std::collections::HashMap<&str, &str> = std::collections::HashMap::new();
149        vars.insert("max_entities", &max_entities_str);
150        vars.insert("max_relations", &max_relations_str);
151        vars.insert("text", text);
152        template
153            .format(&vars)
154            .unwrap_or_else(|_| EXTRACTION_PROMPT.to_string())
155    };
156
157    extract_with_retry(llm, text, prompt).await
158}
159
160/// P2-1/P2-2: 优先原生 tool_calls 结构化输出;文本 JSON 解析失败时,
161/// 带"只返回 JSON"提示重试 1-2 次,仍失败才返回错误。
162async fn extract_with_retry<M: BaseChatModel>(
163    llm: &M,
164    original_text: &str,
165    prompt: String,
166) -> Result<ExtractionResult, super::GraphRAGError> {
167    const MAX_RETRIES: usize = 2;
168    let mut current_prompt = prompt;
169
170    for attempt in 0..=MAX_RETRIES {
171        let result = chat_structured(
172            llm,
173            Some(extraction_tool()),
174            vec![Message::human(&current_prompt)],
175        )
176        .await
177        .map_err(|e| super::GraphRAGError::LLMError(e.to_string()))?;
178
179        if let Some(parsed) = parse_structured(&result) {
180            return Ok(parsed);
181        }
182
183        if attempt < MAX_RETRIES {
184            current_prompt = format!(
185                "上次的输出不是合法 JSON,无法解析。请重新从下面文本提取实体与关系,\
186                 只返回一个 JSON 对象(键为 entities 与 relations),不要包含任何解释、\
187                 编号、引号或代码块。\n\n文本:\n{}\n\n上次输出(无效):\n{}\n\n只输出 JSON 对象:",
188                original_text, result.content
189            );
190        }
191    }
192
193    Err(super::GraphRAGError::ExtractionError(
194        "LLM repeatedly returned invalid JSON; entity/relation extraction failed".to_string(),
195    ))
196}
197
198/// Parses the LLM JSON response into an `ExtractionResult`.
199pub fn parse_extraction(raw: &str) -> Result<ExtractionResult, super::GraphRAGError> {
200    lc_core::json_parse::parse_llm_json::<ExtractionResult>(raw).map_err(|e| {
201        super::GraphRAGError::ExtractionError(format!("Failed to parse extraction JSON: {}", e))
202    })
203}
204
205/// Attempts to extract a JSON object from text that may contain markdown fences.
206#[cfg(test)]
207fn extract_json(text: &str) -> String {
208    let trimmed = text.trim();
209
210    // Case 1: wrapped in ```json ... ```
211    if let Some(rest) = trimmed.strip_prefix("```json") {
212        if let Some(end) = rest.find("```") {
213            return rest[..end].trim().to_string();
214        }
215    }
216
217    // Case 2: wrapped in ``` ... ```
218    if let Some(rest) = trimmed.strip_prefix("```") {
219        if let Some(end) = rest.find("```") {
220            return rest[..end].trim().to_string();
221        }
222    }
223
224    // Case 3: find first { ... last }
225    if let Some(start) = trimmed.find('{') {
226        if let Some(end) = trimmed.rfind('}') {
227            if end > start {
228                return trimmed[start..=end].to_string();
229            }
230        }
231    }
232
233    trimmed.to_string()
234}
235
236#[cfg(test)]
237mod tests {
238    use super::*;
239
240    #[test]
241    fn test_parse_extraction_valid() {
242        let raw = r#"{"entities":[{"name":"Alice","type":"Person","description":"A developer"}],"relations":[{"source":"Alice","target":"Rust","type":"uses","description":"Alice uses Rust"}]}"#;
243        let result = parse_extraction(raw).unwrap();
244        assert_eq!(result.entities.len(), 1);
245        assert_eq!(result.entities[0].name, "Alice");
246        assert_eq!(result.relations.len(), 1);
247        assert_eq!(result.relations[0].source, "Alice");
248    }
249
250    #[test]
251    fn test_parse_extraction_markdown_wrapped() {
252        let raw = r#"```json
253{"entities":[{"name":"Bob","type":"Person","description":"A manager"}],"relations":[]}
254```"#;
255        let result = parse_extraction(raw).unwrap();
256        assert_eq!(result.entities.len(), 1);
257        assert_eq!(result.entities[0].name, "Bob");
258    }
259
260    #[test]
261    fn test_parse_extraction_empty_arrays() {
262        let raw = r#"{"entities":[],"relations":[]}"#;
263        let result = parse_extraction(raw).unwrap();
264        assert!(result.entities.is_empty());
265        assert!(result.relations.is_empty());
266    }
267
268    #[test]
269    fn test_parse_extraction_invalid() {
270        let raw = "not json at all";
271        assert!(parse_extraction(raw).is_err());
272    }
273
274    #[test]
275    fn test_extract_json_plain() {
276        let input = r#"{"key": "value"}"#;
277        assert_eq!(extract_json(input), input);
278    }
279
280    #[test]
281    fn test_extract_json_code_fence() {
282        let input = "```json\n{\"key\": \"value\"}\n```";
283        assert_eq!(extract_json(input), "{\"key\": \"value\"}");
284    }
285
286    #[test]
287    fn test_extract_json_with_surrounding_text() {
288        let input = "Here is the result:\n{\"key\": \"value\"}\nDone.";
289        assert_eq!(extract_json(input), "{\"key\": \"value\"}");
290    }
291
292    /// Verify the extraction prompt contains a few-shot example.
293    #[test]
294    fn test_extraction_prompt_contains_few_shot_example() {
295        assert!(
296            EXTRACTION_PROMPT.contains("Example:"),
297            "extraction prompt should contain few-shot example"
298        );
299        assert!(
300            EXTRACTION_PROMPT.contains("Alice"),
301            "extraction prompt example should contain entity 'Alice'"
302        );
303        assert!(
304            EXTRACTION_PROMPT.contains("works_at"),
305            "extraction prompt example should contain relation type 'works_at'"
306        );
307    }
308
309    /// P2-1: 提取工具定义携带完整 JSON Schema(entities + relations)。
310    #[test]
311    fn test_extraction_tool_schema() {
312        let tool = extraction_tool();
313        assert_eq!(tool.function.name, "extract_entities_relations");
314        let params = tool.function.parameters.expect("parameters should exist");
315        assert!(params["properties"]["entities"].is_object());
316        assert!(params["properties"]["relations"].is_object());
317    }
318
319    /// P2-1: tool_calls 参数可解析为 ExtractionResult。
320    #[test]
321    fn test_parse_structured_tool_args() {
322        let result = StructuredChatResult {
323            content: "".to_string(),
324            tool_args: Some(json!({
325                "entities": [{"name": "Alice", "type": "Person", "description": "dev"}],
326                "relations": []
327            })),
328        };
329        let parsed = parse_structured(&result).expect("tool_args should parse successfully");
330        assert_eq!(parsed.entities.len(), 1);
331        assert_eq!(parsed.entities[0].name, "Alice");
332        assert!(parsed.relations.is_empty());
333    }
334
335    /// P2-1: 无 tool_calls 时回落文本 JSON 解析。
336    #[test]
337    fn test_parse_structured_text_fallback() {
338        let result = StructuredChatResult {
339            content: r#"{"entities": [{"name": "Bob", "type": "Person", "description": "mgr"}], "relations": []}"#.to_string(),
340            tool_args: None,
341        };
342        let parsed = parse_structured(&result).expect("text JSON should parse successfully");
343        assert_eq!(parsed.entities[0].name, "Bob");
344    }
345
346    /// P2-1: tool_args 反序列化失败且文本非 JSON → None(触发重试)。
347    #[test]
348    fn test_parse_structured_none() {
349        let result = StructuredChatResult {
350            content: "not json".to_string(),
351            tool_args: Some(json!("not an object")),
352        };
353        assert!(parse_structured(&result).is_none());
354    }
355}
lc_rag/graph_rag/extractor.rs

lc_rag/graph_rag/
extractor.rs
