Skip to content

module_reference.md

Latest commit

 

History

History
169 lines (121 loc) · 7.34 KB

module_reference.md

File metadata and controls

169 lines (121 loc) · 7.34 KB

Core Module Reference

Pipeline Entry — alg/src/pipeline/functions.py

The main pipeline orchestrates the full answer generation flow through two primary functions:

Function Description
supply_question(body, headers) Phase 1: Receives user query, performs intent understanding (rejection check, supplementary question generation, query translation), and initializes the QA context.
answer(body, headers) Phase 2: Executes the full RAG pipeline — query decomposition, retrieval, answer generation, and timeline construction — returning a streaming response.

Execution Flow of answer():

1. Load QA context from Elasticsearch
2. [Parallel] Query decomposition (GQD) + Retrieval range detection
3. Graph-based multi-source retrieval
4. [Parallel] Answer generation (TAG, streaming) + Timeline generation
5. Further question recommendation
6. Stream results via Server-Sent Events

Query Preprocessing and Intent Understanding

Query Preprocessing: A fine-tuned LLM (Qwen-2.5-14B) filters unsafe/harmful queries and clarifies ambiguous ones by prompting the user with options. A second LLM (Qwen-2.5-14B) rewrites the query by resolving relative spatiotemporal terms (e.g., "last week" → precise timestamp ranges), vague locations ("nearby" → specific places), and incomplete entity names.

Class: IntentionUnderstandingTask — modules/intention_understanding_group/

Method Output
get_intention() Returns rejection decision, supplementary question suggestions, related events, and translated query
get_reject_judgement() Returns only the rejection decision for unsafe/inappropriate content

Output structure:

{
  "question_rejection": {"is_reject": false, "reject_reason": ""},
  "question_supplement": {"is_supply": false, "supply_description": "", "supply_choices": []},
  "related_event": ""
}

Query Decomposition (GQD) — modules/query_division_based_cot_group/

Class: QueryDivisionBasedCoTTask

Decomposes complex queries into a directed acyclic graph (DAG) of sub-queries with explicit dependencies.

Method Description
get_cot(use_scene="general") Entry point; dispatches to query_rewrite_general() or query_rewrite_timeline() based on scene
query_rewrite_general() Decomposes general queries via MultiHopSplitQueries, constructs DAG using IGraph
query_rewrite_timeline() Decomposes timeline-specific queries via TimeLineSplitQueries

DAG Structure (IGraph):

class IGraph:
    """Directed acyclic graph for structured query decomposition."""
    add_new_node(node: ArcNode)          # Add sub-query node
    insert_node_front(new_node, ori_val) # Insert dependency before node
    add_arrow(query, former_val)         # Add dependency edge
    get_turns()                          # BFS execution order respecting dependencies
    get_attr(attr_name)                  # Filter nodes by attribute (need_rag, FunctionCall, etc.)

Each ArcNode contains: sub-query text, need_rag flag, FunctionCall indicator, retrieved answer, and dependency links.

Multi-Source Retrieval — modules/recall_group/

Class: RecallTask

Performs graph-guided multi-source retrieval following the DAG execution order.

Method Description
get_graph_recall(graph, application, retrieval_field, top_n_indices) Traverses the DAG, retrieves documents for each sub-query in parallel
_router(search_field, origin_query) Routes queries to appropriate data sources (news, hot news, general)
_get_recall_queries(graph, application) Extracts retrieval queries from DAG nodes

Retrieval Sources:

  • Online web search (via IAAR Database API) — multiple search engines queried simultaneously
  • Elasticsearch full-text search
  • Milvus vector similarity search

Chunking: Uses RecursiveCharacterTextSplitter with chunk size 350 and overlap 25%, following Azure AI Search findings.

Chunk Deduplication: Computes pairwise cosine similarity using fine-tuned bge-large embeddings. Chunks with similarity > 0.8 are deduplicated via a greedy maximum independent set strategy.

Reranking: Uses fine-tuned bge-reranker-v2-m3 to sort chunks by relevance to each sub-query. Configuration from CommonConfig:

Parameter Default Description
topk_es 1000 Top-K candidates from Elasticsearch
topk_vec 500 Top-K candidates from vector search
topk_rerank 150 Final top-K after reranking

Answer Generation (TAG) — modules/query_answer_group/

Class: QueryAnswerTask

Generates coherent answers by aligning extracted relation triplets with evidence chunks.

Method Description
get_query_answer(query, qa_series_id, ..., streaming=True) Streaming answer generation with inline citations and image placement
get_query_answer_without_streaming(...) Synchronous answer generation

Textual-Visual Choreography: Images from retrieved documents are filtered and placed alongside generated paragraphs:

Step Method Threshold / Model
Image quality filtering Rule-based Removes logos, icons, low-res images
Image relevance filtering bge-reranker-v2-m3 similarity ≥ 0.3
Image-text similarity (1) clip-vit-huge-patch14 paragraph ↔ image embedding
Image-text similarity (2) bge-reranker-v2-m3 paragraph ↔ document title
Image-text similarity (3) bge-large paragraph ↔ document text
Optimal alignment Hungarian algorithm weighted average of (1)(2)(3)

Streaming Output Types:

Type Content
state Pipeline status updates
text Answer text chunk
image Image reference to insert
ref_answer Citation/reference information
text_end End-of-answer signal
recommendation Suggested follow-up questions

Timeline Generation — modules/timeline_group/

Class: TimelineTask

Generates structured timeline visualizations for temporal queries.

Pipeline Steps:

Query Rewrite → Intent Understanding → CoT Query Splitting
→ Multi-source Retrieval → Event Extraction → Deduplication
→ Event Grouping → Highlight Extraction → Granularity Determination → Reference Extraction

Key Parameters:

Parameter Value
Event extraction LLM Qwen2.5-14B-Instruct
Event embedding model bge-large
Event deduplication threshold cosine similarity > 0.9

Runs in parallel with answer generation when pro_flag=True.

Context Management — include/context/RagQAContext.py

Class: RagQAContext

Central context object shared across all pipeline modules within a single QA session.

Category Key Methods
Question add_single_question(request_id, question_id, question, pro_flag), get_single_question()
DAG set_dag(dag), get_dag()
References add_references_result(references), get_references_result_doc(need_new_content)
Serialization context_encode(obj) → Base64 string, context_decode(str) → RagQAContext

Context is persisted in Elasticsearch between the supply_question and answer phases, enabling stateful multi-turn conversations.