feat(kg): add markdown export format for knowledge graph

- Add kg_ocr/export/markdown_exporter.py with export_markdown()
- Cluster entity co-occurrences using Louvain community detection
- Wire -f markdown into `ocr-pipeline kg export` CLI command with KG_REPORT.md default
- Add unit and CLI tests in tests/test_graph.py
- Document markdown export in README.md
- Untrack pycache binaries and ignore graphify-out and KG_REPORT.md

kg_ocr/export/markdown_exporter.py

@@ -0,0 +1,270 @@
+"""Export a knowledge graph to a Markdown report."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from kg_ocr.graph.analyzer import (
+    Anomaly,
+    detect_anomalies,
+    summary,
+    top_citations,
+    top_cooccurrences,
+    top_entities,
+)
+
+
+def _load_graph(graph_input: Any) -> Any:
+    """Load a NetworkX graph from a file path or return it directly."""
+    import networkx as nx
+
+    if isinstance(graph_input, (str, Path)):
+        path = Path(graph_input)
+        data = json.loads(path.read_text(encoding="utf-8"))
+        try:
+            return nx.node_link_graph(data, edges="edges")
+        except TypeError:
+            return nx.node_link_graph(data)
+    return graph_input
+
+
+def _cluster_entity_cooccurrences(graph: Any) -> list[list[dict[str, Any]]]:
+    """Find concept clusters over the entity co-occurrence subgraph."""
+    import networkx as nx
+
+    cooccur_graph = nx.Graph()
+    for node, data in graph.nodes(data=True):
+        if data.get("kind") == "entity":
+            cooccur_graph.add_node(
+                node,
+                text=data.get("text", node.replace("entity:", "")),
+                mentions=int(data.get("mentions", 0)),
+            )
+
+    for u, v, data in graph.edges(data=True):
+        if data.get("kind") == "CO_OCCURS":
+            cooccur_graph.add_edge(u, v, weight=data.get("weight", 1))
+
+    if cooccur_graph.number_of_nodes() == 0:
+        return []
+
+    # Community detection
+    communities: list[set[str]] = []
+    if cooccur_graph.number_of_edges() > 0:
+        try:
+            import networkx.algorithms.community as nx_comm  # type: ignore[import-untyped]
+
+            communities = list(
+                nx_comm.louvain_communities(cooccur_graph, weight="weight", seed=42)
+            )
+        except Exception:
+            try:
+                import networkx.algorithms.community as nx_comm  # type: ignore[import-untyped]
+
+                communities = list(
+                    nx_comm.greedy_modularity_communities(cooccur_graph, weight="weight")
+                )
+            except Exception:
+                try:
+                    import community as community_louvain  # type: ignore[import-not-found]
+
+                    partition = community_louvain.best_partition(
+                        cooccur_graph, weight="weight", random_state=42
+                    )
+                    comm_dict: dict[int, set[str]] = {}
+                    for n, cid in partition.items():
+                        comm_dict.setdefault(cid, set()).add(n)
+                    communities = list(comm_dict.values())
+                except Exception:
+                    communities = list(nx.connected_components(cooccur_graph))
+    else:
+        # No edges: each entity is an isolated node
+        communities = [{n} for n in cooccur_graph.nodes()]
+
+    # Format and sort clusters
+    clusters: list[list[dict[str, Any]]] = []
+    for comm in communities:
+        cluster_nodes = []
+        for node in comm:
+            cluster_nodes.append(
+                {
+                    "id": node,
+                    "text": cooccur_graph.nodes[node].get("text", node),
+                    "mentions": cooccur_graph.nodes[node].get("mentions", 0),
+                }
+            )
+        # Sort entities inside cluster by mentions descending
+        cluster_nodes.sort(key=lambda item: item["mentions"], reverse=True)
+        clusters.append(cluster_nodes)
+
+    # Sort clusters by size (number of entities) descending, then top entity mentions
+    clusters.sort(
+        key=lambda c: (len(c), c[0]["mentions"] if c else 0),
+        reverse=True,
+    )
+    return clusters
+
+
+def export_markdown(graph_path: str | Path, output_path: str | Path) -> Path:
+    """Render a knowledge graph into a Markdown summary report.
+
+    Parameters
+    ----------
+    graph_path : str | Path | nx.Graph
+        Path to `kg_graph.json` or an in-memory NetworkX graph.
+    output_path : str | Path
+        Target path for the exported Markdown report.
+
+    Returns
+    -------
+    Path
+        The output Path written to.
+    """
+    graph = _load_graph(graph_path)
+    out = Path(output_path)
+    out.parent.mkdir(parents=True, exist_ok=True)
+
+    counts = summary(graph)
+    doc_count = counts.get("nodes:document", 0)
+    chunk_count = counts.get("nodes:chunk", 0)
+    entity_count = counts.get("nodes:entity", 0)
+    citation_count = counts.get("nodes:citation", 0)
+    contains_edges = counts.get("edges:CONTAINS", 0)
+    mentions_edges = counts.get("edges:MENTIONS", 0)
+    cites_edges = counts.get("edges:CITES", 0)
+    cooccurs_edges = counts.get("edges:CO_OCCURS", 0)
+
+    entities = top_entities(graph, limit=20)
+    cooccurrences = top_cooccurrences(graph, limit=20)
+    citations = top_citations(graph, limit=20)
+    clusters = _cluster_entity_cooccurrences(graph)
+
+    # Anomalies detection
+    anomalies: list[Anomaly] = list(detect_anomalies(graph))
+    # Detect isolated citations (citations with 0 citing chunks)
+    existing_anomaly_nodes = {a.node for a in anomalies}
+    for node, data in graph.nodes(data=True):
+        if data.get("kind") == "citation" and node not in existing_anomaly_nodes:
+            citing_count = sum(
+                1 for _, _, edge in graph.in_edges(node, data=True) if edge.get("kind") == "CITES"
+            )
+            if citing_count == 0:
+                ident = data.get("identifier", node)
+                anomalies.append(
+                    Anomaly(
+                        "isolated_citation",
+                        node,
+                        f"citation {ident} has no referencing chunks",
+                    )
+                )
+
+    lines: list[str] = [
+        "# Knowledge Graph Report",
+        "",
+        "## Summary",
+        f"- **Documents**: {doc_count:,}",
+        f"- **Chunks**: {chunk_count:,}",
+        f"- **Entities**: {entity_count:,}",
+        f"- **Citations**: {citation_count:,}",
+        f"- **Relationships**: {contains_edges + mentions_edges + cites_edges + cooccurs_edges:,} "
+        f"(CONTAINS: {contains_edges:,}, MENTIONS: {mentions_edges:,}, CITES: {cites_edges:,}, CO_OCCURS: {cooccurs_edges:,})",
+        "",
+    ]
+
+    # Section: Top Entities
+    lines.append("## Top Entities")
+    lines.append("")
+    if entities:
+        lines.append("| Entity | Mentions |")
+        lines.append("|:-------|:---------|")
+        for text, mention_count in entities:
+            clean_text = text.replace("|", "\\|")
+            lines.append(f"| {clean_text} | {mention_count:,} |")
+    else:
+        lines.append("No entities found.")
+    lines.append("")
+
+    # Section: Top Co-occurrences
+    lines.append("## Top Co-occurrences")
+    lines.append("")
+    if cooccurrences:
+        lines.append("| Entity A | Entity B | Co-occurrence Weight |")
+        lines.append("|:---------|:---------|:---------------------|")
+        for left, right, weight in cooccurrences:
+            clean_left = left.replace("|", "\\|")
+            clean_right = right.replace("|", "\\|")
+            lines.append(f"| {clean_left} | {clean_right} | {weight:,} |")
+    else:
+        lines.append("No entity co-occurrences found.")
+    lines.append("")
+
+    # Section: Top Citations
+    lines.append("## Top Citations")
+    lines.append("")
+    if citations:
+        lines.append("| Type | Identifier | Citing Chunks |")
+        lines.append("|:-----|:-----------|:--------------|")
+        for ctype, identifier, count in citations:
+            lines.append(f"| {ctype.upper() or 'UNKNOWN'} | `{identifier}` | {count:,} |")
+    else:
+        lines.append("No citations found.")
+    lines.append("")
+
+    # Section: Concept Clusters
+    lines.append("## Concept Clusters")
+    lines.append("")
+    if clusters:
+        multi_entity_clusters = [c for c in clusters if len(c) > 1]
+        single_entity_clusters = [c for c in clusters if len(c) == 1]
+
+        if multi_entity_clusters:
+            for idx, cluster in enumerate(multi_entity_clusters, start=1):
+                entity_names = [e["text"] for e in cluster[:3]]
+                cluster_label = ", ".join(entity_names)
+                if len(cluster) > 3:
+                    cluster_label += f" (+{len(cluster) - 3} more)"
+                lines.append(f"### Cluster {idx}: {cluster_label}")
+                lines.append(f"**Size**: {len(cluster)} entities")
+                lines.append("")
+                lines.append("| Entity | Mentions |")
+                lines.append("|:-------|:---------|")
+                for item in cluster:
+                    clean_name = item["text"].replace("|", "\\|")
+                    lines.append(f"| {clean_name} | {item['mentions']:,} |")
+                lines.append("")
+
+        if single_entity_clusters:
+            lines.append("### Isolated / Unclustered Entities")
+            lines.append(
+                f"*{len(single_entity_clusters)} entities with no co-occurrences across documents:*"
+            )
+            lines.append("")
+            items_preview = [
+                f"{c[0]['text']} ({c[0]['mentions']})" for c in single_entity_clusters[:30]
+            ]
+            lines.append(", ".join(items_preview))
+            if len(single_entity_clusters) > 30:
+                lines.append(f"*(and {len(single_entity_clusters) - 30} more)*")
+            lines.append("")
+    else:
+        lines.append("No concept clusters detected.")
+        lines.append("")
+
+    # Section: Anomalies
+    lines.append("## Anomalies")
+    lines.append("")
+    if anomalies:
+        lines.append("| Anomaly Type | Node | Detail |")
+        lines.append("|:-------------|:-----|:-------|")
+        for a in anomalies:
+            clean_node = a.node.replace("|", "\\|")
+            clean_detail = a.detail.replace("|", "\\|")
+            lines.append(f"| `{a.kind}` | `{clean_node}` | {clean_detail} |")
+    else:
+        lines.append("No anomalies detected.")
+    lines.append("")
+
+    out.write_text("\n".join(lines), encoding="utf-8")
+    return out