feat: integrate structured extraction and multimodal role-based pipelineFeat/multimodal pipeline

[danielaskdd]

Description

This PR integrates a complete multimodal document pipeline with role-based LLM routing into LightRAG, building on the earlier JSON structured extraction work. It replaces synchronous API-layer document extraction with a proper three-stage pipeline, adds per-role model isolation for all four LLM roles (extract/keyword/query/VLM), and implements full DOCX multimodal support including image extraction, paragraph position tracking, and structured interchange format.

Changes Made

1. JSON Structured Entity Extraction

• Replace delimiter-based extraction with JSON structured output for improved robustness
• Support native JSON mode for OpenAI (response_format: json_object), Ollama (format="json"), and Gemini (response_mime_type="application/json")
• Provider fallback logic when native JSON formatting is unsupported
• Backward compatibility via ENTITY_EXTRACTION_USE_JSON (default: true)
• Auto-detect JSON vs delimiter format during cache rebuild

2. Three-Stage Multimodal Document Pipeline

• PARSE: Structured DOCX extraction with heading-aware semantic chunking, table structure preservation, and OMML formula extraction
• ANALYZE: VLM-based multimodal analysis with sidecar writeback for drawings, tables, and equations
• PROCESS: Entity/relation extraction from text and multimodal chunks, graph and vector store construction
• Pipeline status tracking: PENDING → PARSING → ANALYZING → PROCESSING → PROCESSED/FAILED
• Parser engine routing via LIGHTRAG_PARSER env var and filename hints (supports native/mineru/docling)

3. DOCX Upload Deferred to Pipeline

• .docx uploads no longer do synchronous content extraction at the API layer
• Files are enqueued with pending_parse format; the pipeline's parse_native() handles structured extraction
• Supports parse_docx_to_interchange_jsonl with heading semantics, para_id positions, and image asset extraction

4. Interchange JSONL Enhancements

• positions field now populated with paraid entries from Word w14:paraId attributes
• engine_capabilities includes "i" when embedded images are detected
• asset_dir flag set when *.blocks.assets directory is created

5. Image Binary Extraction & Assets Directory

• New extract_docx_images() extracts embedded images via doc.part.rels relationship API
• Images written to *.blocks.assets/ directory alongside interchange JSONL
• _extract_drawing_info() extended to return r:embed relationship ID
• Paragraph dataclass extended with drawing_rIds for image-paragraph association

6. Role-Based LLM/VLM Routing

• Four independent roles: extract, keyword, query, vlm — each with its own function, concurrency queue, timeout, and provider options
• Runtime reconfiguration via update_llm_role_config() with atomic rollback on failure
• Per-role provider option overrides via options_dict_for_role() (e.g., EXTRACT_OPENAI_LLM_TEMPERATURE)
• Ollama role functions now explicitly bind role_model instead of falling back to global config
• Cross-provider kwargs isolation: base ollama kwargs won't pollute openai role calls
• Ollama API /generate and /chat bypass paths use query_llm_model_kwargs

7. Relation Merge Robustness

• Defensive timeout handling for relation VDB upserts
• Fine-grained logging around entity/relation upsert stages
• Improved observability for edge-processing waits and pending tasks

Test Results

Offline Test Suite

315 passed, 1 skipped, 0 failed (48.97s)

End-to-End Test with Real Multimodal DOCX

Tested with Chapter 2 of a real academic paper (79 paragraphs, 6 tables, 6 embedded images, 25 OMML formulas)：

Verification Point	Result
extraction_format	interchange_jsonl (not legacy)
format_version	2.0
engine_capabilities	["t", "i"] (tables + images)
positions field	Contains paraid entries (e.g., ["1D6D83BF", "68ED75D1"])
*.blocks.assets directory	Created with 6 extracted PNG images
Formula Q&A	Correctly explained formulas (2-1) and (2-2) for self-attention
Table Q&A	Correctly listed 8 comparison features from Table 2-1
Image Q&A	Correctly described all 6 figures with structural details
Image detail Q&A	Described CLIP model diagram components: dual encoders, N×N similarity matrix, arrows, color scheme

Sample Q&A: Image Structure Details

Question: "图2-1中具体包含了哪些视觉元素和组成部分？请详细描述这张图的结构布局"

Answer (abridged):

图2-1 以"CLIP 模型结构与对比学习"为主题，整体采用上下分区、左右对称的布局。上半部分为模型结构（左侧"图像编码器" + 右侧"文本编码器"双塔），下半部分为 N×N 相似度矩阵与对称交叉熵损失。垂直粗箭头从模型结构区指向矩阵，标注"计算余弦相似度"；循环虚线箭头标注"反向传播优化"形成训练闭环。图像侧使用蓝色调，文本侧使用绿色调...

Related Issues

• Supersedes closed PR feat: Entity extraction uses JSON structured output instead of delimiter-based text #2684

Checklist

• Changes tested locally
• Pre-commit checks pass (ruff-format, ruff, trailing-whitespace, end-of-file, requirements-txt-fixer)
• 315 offline tests pass
• End-to-end multimodal DOCX test verified (upload → parse → extract → query)
• Unit tests added for role isolation, runtime updates, rollback, provider options, Ollama kwargs