feat: speed up with parallel call and support for vLLM deployment emb…

[wuxuedaifu]

PR: Parallel Execution & Local Inference Optimization(2X+ faster)

🚨 Existing Problems

1. Synchronous Data Ingestion: Knowledge Graph text portions were ingested sequentially, causing severe performance degradation scaling alongside local LLM calls.
2. I/O Thread Blocking: Profile generation locked the main thread entirely to write tracking logs (CSV/JSON) per entity, creating quadratic O(N²) delays.
3. Infinite Generation Deadlocks (vLLM): Using strict json_object endpoints without max_tokens limits frequently caused local models to hallucinate infinite whitespace loops, permanently hanging execution threads.
4. Inefficient ReACT Loops: The Report Agent rigorously enforced min_tool_calls=3, forcing aware agents to blindly search tools redundantly.
5. Hardcoded Scaling Limits: Core concurrency modifiers (like thread counts and batch sizes) were hardcoded across various backend functions, preventing user-specific deployment tuning.
6. Inflexible Embeddings: The embedding services lacked seamless configuration to support bare-metal or containerized local inference integrations (Ollama / vLLM).

---

🛠️ Key Improvements & Solutions

1. Parallel Knowledge Graph Pipeline (graph_builder.py, neo4j_storage.py)

• Replaced iterative chunking mechanisms inside add_text_batch natively utilizing a ThreadPoolExecutor.
• Concurrent graph interactions safely manage transactional Neo4j MERGE deadlocks via the existing TransientError exponential backoff implementation.

2. High-Throughput Profile Generation (oasis_profile_generator.py)

• Background I/O Pipelines: Replaced sequential main-thread file persistence with threading.Thread fire-and-forget daemon tasks.
• Concurrent DB Queries: Split sequential edges and nodes hybrid searches into simultaneous tandem DB queries, halving query wait times.
• Anti-Deadlocking Constraints: Secured completions endpoints assigning max_tokens=4000. Combined directly with the _fix_truncated_json parser, the system automatically salvages any trailing JSON syntax instead of permanently hanging 128k context windows.

3. Streamlined Report Processing (report_agent.py)

• Standardized report subsections to build entirely asynchronously minimizing total payload generation times.
• Stripped the min_tool_calls metric allowing context-sufficient agents to generate responses instantly.

4. Dynamic Execution Metrics (config.py, .env)

• Centralized hardware scaling configurations up to .env targeting adaptable performance across deployments natively:
  • GRAPH_BUILD_BATCH_SIZE=10
  • PROFILE_PARALLEL_COUNT=10
  • PROFILE_SEARCH_WORKERS=2
  • REPORT_PARALLEL_SECTIONS=5

5. Extended Local Deployment Support (embedding_service.py)

• Abstracted embedding initialization parameters fully restoring unified connectivity targets mapping generalized configurations standardizing Ollama/vLLM accessibility automatically.