A Research-Grade Memory-Augmented LLM Agent for Academic Literature Analysis
面向学术科研文献的记忆增强 LLM Agent — LangGraph · MultiQuery · Hybrid RAG · Persistent Memory
A production-grade AI agent that combines retrieval-augmented generation, multi-query hybrid search, persistent three-tier memory, and automated survey writing into a unified CLI tool for academic research.
# Index 100 arXiv papers in seconds (deduplication built-in)
You > bulk_search RAG retrieval augmented generation LLM 100
[✓ 100 papers indexed | duplicates automatically skipped]
# Questions in ANY language — agent queries tools in English automatically
You > RAG系统最新的检索优化方法有哪些？
Agent → [tool call: "recent retrieval optimization methods RAG 2024"]
Agent > 根据最新论文，主要方法包括：(1) MultiQueryRetrieval...
来源: [CRAG, arXiv:2401.15884] [Self-RAG, arXiv:2310.11511]
# Save research insights to permanent memory
You > remember HyDE: embed hypothetical answer instead of query for better recall
[✓ Saved to memory #1 — injected into system prompt next startup]
# Generate a full survey (grounded in indexed papers)
You > survey Retrieval-Augmented Generation for Large Language Models
[Writing 8 sections via Map-Reduce, each grounded by hybrid retrieval...]
[✓ Saved: survey_Retrieval-Augmented_Generation_20260417_1800.md]
| Feature | Implementation | Research Basis |
|---|---|---|
| MultiQuery Retrieval | LLM generates 3 query variants → merge & dedup | Ma et al. (2023) — +20-40% recall |
| Hybrid Search | BM25 (sparse) + FAISS MMR (dense) + RRF fusion | Cormack et al., SIGIR 2009 |
| LangGraph Agent | create_react_agent + MemorySaver |
Replaces deprecated AgentExecutor |
| FastEmbed | BAAI/bge-small-en-v1.5 via ONNX Runtime |
No PyTorch — ~150 MB vs ~1 GB |
| Chain-of-Thought Prompt | 5-step reasoning + citation guidelines | Wei et al., NeurIPS 2022 |
| arXiv Deduplication | Persistent ID set across sessions | Prevents vector store bloat |
| Bulk Indexing | arXiv API → 100+ paper abstracts | Rate-limited, progress bar |
| Survey Writer | Map-Reduce pipeline, 8 sections | RAG-grounded, saves to .md |
| 3-Tier Memory | Working / Episodic / Semantic | Tulving (1972); Park et al. (2023) |
| Free API | SiliconFlow Qwen / Groq Llama | Zero cost to start |
The full retrieval stack for every query:
User question (any language)
│
▼ [translated to English by Agent]
MultiQueryRetriever ─── generates 3 query variants via LLM
│ e.g. "RAG" → ["retrieval augmented generation",
│ "dense retrieval LLM grounding",
│ "knowledge-grounded language model"]
▼
For each variant:
┌── BM25 keyword search → ranked list A
└── FAISS MMR dense search → ranked list B
│
▼
RRF Fusion score = Σ 1 / (60 + rank + 1)
│
▼
Deduplicated Top-K results
- BM25 — exact keyword match, best for technical terms & author names
- FAISS MMR — semantic similarity with diversity (prevents redundant results)
- RRF — rank fusion without score calibration (
k=60, Cormack et al. SIGIR 2009) - MultiQuery — 3 phrasings per query, published to improve recall by 20–40%
| Tier | Type | Scope | Commands |
|---|---|---|---|
| 1 — Working | LangGraph MemorySaver |
Current session | automatic |
| 2 — Episodic | Session archive (JSON) | Permanent, last 10 | clear_memory, exit |
| 3 — Semantic | User fact notes (JSON) | Permanent | remember / forget |
Cross-session memory injection: On every startup, Tier 2 & 3 facts are injected into the agent's system prompt — the agent "remembers" your past research even after restart.
Inspired by Tulving (1972) Episodic & Semantic Memory, Park et al. (2023) Generative Agents, Zhong et al. (2024) MemoryBank
Map-Reduce workflow: for each of the 8 sections, the MultiQuery+Hybrid retriever finds the most relevant paper chunks; the LLM writes that section grounded in retrieved context; all sections are assembled into a final Markdown document.
Output sections: Abstract · Introduction · Background · Taxonomy · Methods · Applications · Challenges & Limitations · Conclusion & Future Directions
git clone https://github.com/YOUR_USERNAME/academic-rag-agent.git
cd academic-rag-agent
conda env create -f environment.yml
conda activate rag-agentcopy .env.example .env
# Edit .env — fill in OPENAI_API_KEY| Provider | Sign Up | Recommended Free Model |
|---|---|---|
| SiliconFlow | Phone / Email | Qwen/Qwen2.5-72B-Instruct |
| Groq | Google account | llama-3.3-70b-versatile |
OPENAI_API_KEY=sk-your-key-here
OPENAI_BASE_URL=https://api.siliconflow.cn/v1
OPENAI_MODEL=Qwen/Qwen2.5-72B-Instruct
EMBEDDING_PROVIDER=local
EMBEDDING_MODEL=BAAI/bge-small-en-v1.5conda activate rag-agent
cd path\to\academic-rag-agent
python main.py| Command | Description |
|---|---|
load <path> |
Index a PDF / TXT / MD file |
load_dir <dir> |
Index all docs in a folder |
| Command | Description |
|---|---|
bulk_search <topic> [N] |
Fetch & index N abstracts from arXiv (default 50, dedup enabled) |
bulk_search retrieval augmented generation survey 100
| Command | Description |
|---|---|
survey <topic> |
Generate a complete 8-section survey paper (Markdown) |
survey <topic> --out file.md |
Save to a specific file |
bulk_search RAG large language model 80
survey Retrieval-Augmented Generation for LLMs
| Command | Description |
|---|---|
remember <note> |
Save a fact to permanent memory (survives restart) |
memories |
Show all saved facts & session archive |
forget <id> |
Delete a memory by ID |
clear_memory |
Auto-archive session summary and reset chat |
| Command | Description |
|---|---|
status |
Show KB size, memory stats, paths |
clear_db |
Wipe the vector store |
help |
Show all commands |
exit |
Auto-save session summary and quit |
academic-rag-agent/
├── main.py # CLI entry point — all commands
├── requirements.txt
├── environment.yml # Conda environment
├── .env.example # Config template (no secrets)
├── docs/
│ └── images/ # Architecture diagrams
└── src/
├── agent/
│ └── agent.py # LangGraph ReAct agent + CoT prompt
├── rag/
│ ├── document_loader.py # PDF/TXT/MD loader with file dedup
│ ├── vector_store.py # FAISS + MultiQuery + FastEmbed + arXiv dedup
│ └── hybrid_retriever.py # BM25 + FAISS + RRF fusion
├── memory/
│ └── persistent_memory.py # Three-tier persistent memory
└── tools/
├── arxiv_tool.py # arXiv live search (LangChain tool)
├── bulk_arxiv.py # Bulk metadata fetcher (100+ papers)
└── survey_writer.py # Map-Reduce survey generation
| Technology | Reference |
|---|---|
| RAG | Lewis et al., RAG for Knowledge-Intensive NLP (NeurIPS 2020) |
| ReAct Agent | Yao et al., ReAct: Synergizing Reasoning and Acting (ICLR 2023) |
| MultiQuery Retrieval | Ma et al., Query Rewriting for RAG (2023) |
| Chain-of-Thought | Wei et al., Chain-of-Thought Prompting (NeurIPS 2022) |
| RRF Fusion | Cormack, Clarke & Buettcher, RRF outperforms Condorcet (SIGIR 2009) |
| MMR Retrieval | Carbonell & Goldstein, The Use of MMR (SIGIR 1998) |
| Memory System | Tulving (1972); Park et al., Generative Agents (2023) |
| LangGraph | LangGraph Docs |
- Hybrid BM25 + FAISS + RRF retrieval
- MultiQueryRetriever (Ma et al. 2023, +20-40% recall)
- arXiv paper deduplication across sessions
- LangGraph
create_react_agentwith Chain-of-Thought prompt - FastEmbed (ONNX, no PyTorch)
- Bulk arXiv indexing (100+ papers)
- Map-Reduce survey generation (8 sections)
- Three-tier persistent memory (Working / Episodic / Semantic)
- Cross-encoder reranker (BGE-reranker, Nogueira et al. 2019)
- HyDE — Hypothetical Document Embeddings (Gao et al. 2022)
- RAGAS automated evaluation (faithfulness / relevancy metrics)
- Streamlit / Gradio web UI
- Multi-agent collaboration (researcher + critic + writer)
MIT License