2nd/10_Wiki/Topics/Other/Knowledge synthesis.md

---
id: wiki-2026-0508-knowledge-synthesis
title: Knowledge Synthesis
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [지식 종합, Synthesis, Information Integration]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [research, knowledge, synthesis, methodology, llm]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: rag
---

# Knowledge Synthesis

## 매 한 줄
> **"매 synthesis = 차이의 통합"**. 여러 source 의 fragmented knowledge 를 단일 coherent representation 으로 통합하는 process. 2026 LLM 시대에 RAG + structured extraction + cross-document reasoning 이 표준 toolchain.

## 매 핵심

### 매 5-step pipeline
1. **Acquisition**: search / scrape / corpus collection.
2. **Extraction**: structured fact / claim 추출.
3. **Alignment**: same entity / concept matching across sources.
4. **Integration**: conflict resolution + gap filling.
5. **Presentation**: report / map / model.

### 매 synthesis types
- **Aggregative**: meta-analysis, systematic review.
- **Interpretive**: thematic synthesis, narrative review.
- **Generative**: 새 framework / theory 제시.

### 매 응용
1. Systematic literature review.
2. Competitive intelligence.
3. Wiki / knowledge graph build-out.
4. Investigation reporting.

## 💻 패턴

### RAG synthesis pipeline
```python
from anthropic import Anthropic
import chromadb

client = Anthropic()
db = chromadb.Client().create_collection("docs")

def synthesize(question: str, k=8):
    docs = db.query(query_texts=[question], n_results=k)["documents"][0]
    context = "\n---\n".join(docs)
    return client.messages.create(
        model="claude-opus-4-7",
        max_tokens=4096,
        system="Synthesize the sources. Cite [N]. Note conflicts.",
        messages=[{"role": "user", "content": f"Q: {question}\nSources:\n{context}"}],
    ).content[0].text
```

### Structured extraction
```python
from pydantic import BaseModel

class Claim(BaseModel):
    statement: str
    source: str
    confidence: float
    contradicts: list[str] = []

def extract_claims(text: str, source: str) -> list[Claim]:
    resp = client.messages.create(
        model="claude-opus-4-7",
        max_tokens=2048,
        system="Extract atomic claims as JSON list of {statement, confidence}.",
        messages=[{"role": "user", "content": text}],
    )
    return [Claim(source=source, **c) for c in json.loads(resp.content[0].text)]
```

### Entity alignment
```python
from sentence_transformers import SentenceTransformer
import numpy as np

model = SentenceTransformer("all-MiniLM-L6-v2")

def align(entities_a, entities_b, threshold=0.85):
    emb_a = model.encode(entities_a)
    emb_b = model.encode(entities_b)
    sim = emb_a @ emb_b.T / (np.linalg.norm(emb_a, axis=1)[:, None] *
                              np.linalg.norm(emb_b, axis=1)[None, :])
    pairs = []
    for i, j in zip(*np.where(sim > threshold)):
        pairs.append((entities_a[i], entities_b[j], sim[i, j]))
    return pairs
```

### Conflict detection
```python
def detect_conflicts(claims: list[Claim]) -> list[tuple]:
    pairs = []
    for i, c1 in enumerate(claims):
        for c2 in claims[i+1:]:
            resp = client.messages.create(
                model="claude-haiku-4-5",
                max_tokens=10,
                messages=[{"role": "user",
                           "content": f"Contradict?\nA: {c1.statement}\nB: {c2.statement}\nYes/No"}],
            )
            if "yes" in resp.content[0].text.lower():
                pairs.append((c1, c2))
    return pairs
```

### Hierarchical summarization
```python
def hier_summarize(docs: list[str], chunk=4) -> str:
    if len(docs) <= 1:
        return docs[0] if docs else ""
    summaries = []
    for i in range(0, len(docs), chunk):
        merged = "\n".join(docs[i:i+chunk])
        s = spawn_summarize(merged)
        summaries.append(s)
    return hier_summarize(summaries, chunk)
```

### Citation graph build
```python
import networkx as nx

def build_graph(claims: list[Claim]) -> nx.DiGraph:
    g = nx.DiGraph()
    for c in claims:
        g.add_node(c.statement, source=c.source, conf=c.confidence)
        for contra in c.contradicts:
            g.add_edge(c.statement, contra, type="contradicts")
    return g
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| 수십 개 sources | RAG + hierarchical summarize |
| Conflicting claims | Structured extraction + LLM judge |
| Novel framework needed | Generative synthesis |
| Quantitative pooling | Meta-analysis stats |

**기본값**: RAG + structured extraction + claim graph.

## 🔗 Graph
- 부모: [[Research Methodology]] · [[Information Retrieval]]
- 응용: [[RAG]] · [[Knowledge Graph]]

## 🤖 LLM 활용
**언제**: literature review, competitive analysis, contradictory source reconciliation.
**언제 X**: domain-novel concept invention without grounding — hallucination 위험.

## ❌ 안티패턴
- **Cherry-picking**: 매 confirming source 만 select.
- **Source-blind synthesis**: citation 없이 claim 통합 → traceability 상실.
- **Premature consensus**: conflict 무시하고 단일 narrative.
- **Hallucinated alignment**: LLM 이 entity 임의 동일시.

## 🧪 검증 / 중복
- Verified (PRISMA 2020, Cochrane Handbook 6.4, Anthropic RAG cookbook).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — 5-step pipeline + RAG/extraction 패턴 |