• 学术相关
  • HippoRAG
    • 0. Execution Overview
    • 1. High-level Design (Indexing → Retrieval → Generation)
    • 2. Offline Construction: Indexing (Detailed Execution)
    • 3. Online Query: Retrieval (Detailed Execution)
    • 4. Online Generation: Generation (Detailed Execution)
    • 5. Key Design Decisions
    • 6. Evaluation
    • 7. Limitations and Applicability
    • 8. Quick Reference
  • HippoRAG2
    • 0. Execution Overview
    • 1. High-level Design (Indexing → Retrieval → Generation)
    • 2. Offline Construction: Indexing (Detailed Execution)
    • 3. Online Query: Retrieval (Detailed Execution)
    • 4. Online Generation: Generation (Detailed Execution)
    • 5. Key Design Decisions
    • 6. Evaluation
    • 7. Limitations and Applicability
    • 8. Quick Reference

学术相关

HippoRAG

One-sentence positioning: A retrieval framework inspired by the hippocampal indexing theory that extracts an open knowledge graph via LLM and combines it with the Personalized PageRank algorithm to achieve cross-passage multi-hop knowledge integration in a single retrieval step.

Key innovation: Uses Personalized PageRank to perform pattern completion on the knowledge graph, compressing traditional iterative multi-hop retrieval into a single-step graph traversal, achieving performance comparable to or better than iterative retrieval while reducing cost by 10-30x and improving speed by 6-13x.

0. Execution Overview

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Offline Phase (one-time construction)
  ├─ ① Named entity recognition (extract named entities from each passage)
  ├─ ② OpenIE extraction of KG triples (two-stage: NER → OpenIE extracts noun phrase nodes N and relation edges E)
  ├─ ③ Retrieval encoder supplements synonym edges (add E' between entities with cosine similarity > τ)
  └─ ④ Construct Passage-Node co-occurrence matrix P (|N| × |P|, recording each noun phrase's occurrence count in each passage)
         ↓
Online Phase (per query)
  ├─ ⑤ Query entity extraction (LLM extracts salient named entities from query)
  ├─ ⑥ Query node mapping (entity vectorized encoding, mapped to most similar KG node via cosine similarity)
  ├─ ⑦ Run Personalized PageRank (query nodes as seeds, PPR distributes probability on graph)
  │      - Initialization: query nodes equal probability, rest are 0
  │      - Node-specific weighting: si = |Pi|^{-1} (similar to TF-IDF)
  │      - Transition probability constructed via adjacency matrix
  └─ ⑧ Aggregate PPR node probabilities with co-occurrence matrix P to obtain passage ranking scores
         ↓
  ⑨ Retrieved passages input to LLM for final answer generation

1. High-level Design (Indexing → Retrieval → Generation)

1.1 Indexing

1.2 Retrieval

1.3 Generation

2. Offline Construction: Indexing (Detailed Execution)

Step 2.1 Named Entity Recognition

Step 2.2 OpenIE Extraction of KG Triples

Step 2.3 Supplement Synonym Edges

Step 2.4 Construct Passage-Node Co-occurrence Matrix

3. Online Query: Retrieval (Detailed Execution)

3.1 Retrieval Procedure

Step 3.1 Query Entity Extraction

• Operation: LLM extracts salient named entities (query named entities) from the query
• Purpose: Transform natural language query into seed nodes on the graph

Step 3.2 Query Node Mapping

• Operation: Vectorize query named entities using the retrieval encoder, map to most similar nodes in the KG based on cosine similarity
• Output: Query nodes

Step 3.3 Personalized PageRank Execution

• Initialize personalized probability distribution: All query nodes have equal probability, other nodes probability is 0

• Node-specific weighting: si =

  Pi

  ^{-1}, similar to TF-IDF’s inverse document frequency idea

• Transition probability: Constructed via adjacency matrix
• Core mechanism: PPR distributes probability on the graph only through user-defined source nodes (query nodes), simulating pattern completion in hippocampal neural pathways

Step 3.4 Passage Ranking

• Operation: Multiply the updated PPR node probability distribution with the co-occurrence matrix P
• Output: Final ranking score for each passage

4. Online Generation: Generation (Detailed Execution)

HippoRAG is positioned as a retrieval framework; the notes do not explicitly document special design in the generation phase. Retrieved and ranked passages serve as context input to the downstream LLM for answer generation.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
User query
    │
    ▼
┌─────────────────┐
│ Query entity    │  ← LLM extracts named entities
│ extraction      │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Query node      │  ← Cosine similarity matches KG nodes
│ mapping         │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ PPR probability │  ← Execute Personalized PageRank on graph
│ diffusion       │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ Passage ranking │  ← PPR probability × co-occurrence matrix P
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ LLM generates   │  ← Retrieved passages as context
│ answer          │
└─────────────────┘

5. Key Design Decisions

6. Evaluation

6.1 Evaluation Metrics

6.2 Comparative Experimental Setup

Benchmark datasets: MuSiQue, 2WikiMultiHopQA, HotpotQA

Key findings:

• Single-step retrieval achieves performance comparable to or better than iterative retrieval IRCoT
• Cost is 1/10 to 1/30 of IRCoT, speed is 6-13x faster than IRCoT
• Integration into IRCoT yields further improvements

7. Limitations and Applicability

Best Applicable Scenarios

• Multi-hop QA tasks requiring cross-passage knowledge integration
• Scenarios sensitive to retrieval latency (single-step PPR is 6-13x faster than iterative retrieval)
• Cost-constrained environments (10-30x cheaper than iterative retrieval)
• Scenarios where query information can primarily be expressed through entity relations

Unsuitable Scenarios

• Scenarios where key information in queries cannot be extracted via NER
• Queries requiring extensive contextual cues rather than entity relations
• Scenarios with extremely high demands on fine-grained semantic differences between concepts

8. Quick Reference

HippoRAG2

One-sentence positioning: An improved version of HippoRAG that introduces passage nodes into the knowledge graph to achieve dense-sparse fusion of concepts and context, combining Query-to-Triple retrieval with LLM filtering to comprehensively outperform standard RAG on factual memory, sense-making, and associative memory tasks.

Key innovation: Introduces passage nodes on top of PPR to achieve dense-sparse fusion of concepts and context, and through the Query-to-Triple + LLM filtering retrieval strategy, addresses the performance degradation of knowledge graph RAG on factual memory tasks.

0. Execution Overview

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Offline Phase (one-time construction)
  ├─ ① Entity extraction
  ├─ ② Extract KG triples (OpenIE)
  ├─ ③ Synonym edge completion (cosine similarity > τ)
  └─ ④ Dense-Sparse fusion (core design)
         - Sparse layer: phrase nodes
         - Dense layer: complete vector representation of each passage abstracted as passage node
         - Bridging: passage nodes connected to corresponding entities via contains edges
         ↓
Online Phase (per query)
  ├─ ⑤ Query embedding
  ├─ ⑥ Vector recall
  │      - top-k passages (embedding similarity)
  │      - top-k KG triples (Query to Triple)
  ├─ ⑦ LLM filters triples (produces T' ⊆ T)
  ├─ ⑧ Seed node selection
  │      - If filtered set is empty → directly return top-k passages
  │      - Otherwise
  │        · Phrase nodes: identified from T', top-k selected by average ranking score, probabilities assigned by normalized ranking scores
  │        · Passage nodes: all recalled passages serve as seeds, probability = embedding similarity × weight factor
  ├─ ⑨ Merge & global normalization
  ├─ ⑩ Execute Personalized PageRank
  └─ ⑪ Sort and output retrieval results
         ↓
  ⑫ Retrieved passages input to LLM for final answer generation

1. High-level Design (Indexing → Retrieval → Generation)

1.1 Indexing

1.2 Retrieval

1.3 Generation

2. Offline Construction: Indexing (Detailed Execution)

Step 2.1 Entity Extraction

Step 2.2 Extract KG Triples

Step 2.3 Synonym Edge Completion

Step 2.4 Dense-Sparse Fusion (Core Design)

Why this approach? Concepts are concise and generalizable but lose information; context is semantically rich but adds complexity. Dense-sparse fusion retains the advantages of both, resolving the concept-context trade-off.

3. Online Query: Retrieval (Detailed Execution)

3.1 Retrieval Mode Overview

HippoRAG 2 supports three query mapping methods, with Query to Triple as the default:

3.2 Retrieval Procedure

Step 3.1 Query Embedding

• Operation: Encode the query as a vector representation

Step 3.2 Vector Recall

• top-k passages: Recalled via embedding similarity
• top-k triples: Query to Triple, recall triples in the graph via embedding similarity

Step 3.3 LLM Filters Triples

• Operation: Use LLM to filter retrieved triples T, producing T’ ⊆ T
• Purpose: Improve retrieval quality, remove irrelevant triples

Step 3.4 Seed Node Selection

• If filtered set is empty: Directly return embedding-recalled top-k passages
• Otherwise:
  • Phrase nodes: Identify phrase nodes from filtered triples T’, select top-k by average ranking score, assign probabilities based on normalized ranking scores
  • Passage nodes: All recalled passage nodes also serve as seed nodes, with initial probability being embedding similarity multiplied by a weight factor (§6.2), balancing the influence of phrase nodes and passage nodes

Step 3.5 Merge and Global Normalization

• Merge probability distributions of phrase nodes and passage nodes
• Global normalization

Step 3.6 Personalized PageRank

• Execute PPR using the merged seed node probability distribution as the personalized vector

Step 3.7 Sort and Output

• PPR output node probabilities multiplied by the co-occurrence matrix to obtain final passage ranking scores

4. Online Generation: Generation (Detailed Execution)

HippoRAG 2 is positioned as a retrieval framework; the notes do not explicitly document special design in the generation phase. Retrieved and ranked passages serve as context input to the downstream LLM for answer generation.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
User query
    │
    ▼
┌─────────────────┐
│ Query embedding │
└────────┬────────┘
         │
    ┌────┴────┐
    ▼         ▼
┌────────┐ ┌─────────────┐
│ top-k  │ │ top-k       │
│passages│ │ triples     │
└───┬────┘ └──────┬──────┘
    │             │
    │             ▼
    │      ┌─────────────┐
    │      │ LLM filters │
    │      └──────┬──────┘
    │             │
    │      ┌──────┴──────┐
    │      ▼             ▼
    │  ┌────────┐   ┌──────────┐
    │  │ T' is  │   │ T' is    │
    │  │ empty  │   │ non-empty│
    │  └───┬────┘   └────┬─────┘
    │      │             │
    ▼      ▼             ▼
┌─────────────────────────────────┐
│ Directly return passages        │   ← T' is empty
│ or                              │
│ Phrase nodes + Passage nodes    │   ← T' is non-empty
│ → merge & normalize → PPR       │
└─────────────────────────────────┘
         │
         ▼
┌─────────────────┐
│ Passage ranking │
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ LLM generates   │
│ answer          │
└─────────────────┘

5. Key Design Decisions

6. Evaluation

6.1 Evaluation Metrics

6.2 Comparative Experimental Setup

Datasets:

Key findings:

• Comprehensively superior to standard RAG on factual memory, sense-making, and associative memory tasks
• 7% improvement over SOTA embedding models on associative memory tasks
• Resolves HippoRAG’s performance degradation on factual memory tasks

7. Limitations and Applicability

Notes do not explicitly document specific limitations. From a design perspective, while the concept-context trade-off is mitigated through fusion, the dense-sparse weight balancing still relies on hyperparameter tuning.

Best Applicable Scenarios

• RAG scenarios requiring simultaneous consideration of factual memory and associative reasoning
• Environments requiring multi-hop QA without bearing the high cost of iterative retrieval
• Scenarios where query information can be expressed through both entity relations and contextual context

Unsuitable Scenarios

• Notes do not explicitly document unsuitable scenarios

8. Quick Reference

• Previous
  【阅】本周阅读摘选2026-05-11 → 2026-05-17
• Next
  【阅】本周阅读摘选2026-05-25 → 2026-05-31