GLYPH

GLYPH builds a knowledge graph from documents and code, then serves graph-aware context for retrieval. Document entities come from LLM extraction, code structure from tree-sitter, both behind one extractor port. Retrieval is benchmarked against a fair vector baseline with bootstrap confidence intervals — across both the document and code domains, with the honest result reported either way.

Part of the AXON stack — GLYPH is the graph-retrieval layer that AXON (the product front) consumes to decide what context to bring; rtkx decides how compact. GLYPH also stands alone as a benchmarked knowledge-graph retrieval library.

Why it exists

Vector similarity retrieval ignores structure: who cites whom, what relates to what. For corpora with rich entities and relationships (rules, technical documents) and for code (calls, imports), a graph delivers context that vectors cannot see. GLYPH builds this graph from both domains under a single abstraction, and measures when the graph beats the vector and when it does not justify the cost.

What it does

• Builds knowledge graphs from documents (LLM extraction) and from code (tree-sitter, deterministic).
• Serves graph-aware retrieval: given a query, anchors entities and expands neighborhood by hops.
• Compares against a fair vector baseline on the same corpus, measured with confidence intervals.

What it does NOT do (yet)

• Does not perform complete cross-language type inference in code. Symbol resolution is by name in the import graph and intra-file. Declared limitation (ADR-G5).
• The retrieval token budget is estimated per character; generation tokens in the benchmark are actual model counts.
• Does not replace AXON parsing; GLYPH is the canonical graph source and AXON delegates to it (dec-116 / ADR-G6).

Current status

Phases 0–7 complete. Model + ports + NetworkX adapter (F0); document extraction by LLM with persisted graph (F1); graph-aware retrieval + fair vector baseline + hybrid under a single contract (F2); benchmark against GNOMON with bootstrap CIs (F3); code extraction by tree-sitter (F4); product boundary GraphContextSource consumed by AXON (F5, ADR-G6); packaging & publication (F6); global community axis (F7, ADR-G7).

Published metrics (validation-first):

• Documents (Monster Manual, n=25): reranked_vector (cross-encoder reranker over the vector baseline) now leads context_precision (0.720, above the 0.55+ target); faithfulness ties at 1.000 across graph_pagerank/vector/hybrid (0.980 for graph/reranked_vector) — no single arm dominates. Generator for this run is openai/gpt-oss-120b (free-tier substitute for Claude Haiku 4.5, declared not assumed; judge unchanged, deepseek/deepseek-chat). METRICS.md.
• Code (AXON graph, n=14, judged by two independent judges, Gemini + Qwen): on faithfulness (robust metric — same ordering across both judges) the vector baseline leads and graph ranks last (vector > hybrid > graph); context_precision is judge-dependent (ordering reverses between judges → no reliable winner). The thesis "graph wins on code" did not hold up on the robust metric. METRICS-code.md (Gemini) + METRICS-code-qwen.md (Qwen).
• Global / sense-making (community axis ADR-G7, n=8, two judges): no arm wins significantly in quality (overlapping CIs), but community summary delivers equal-or-better quality at ~half the tokens (5.3k vs 10k+) of vector/graph — same answer, half the cost for "how does this organize?" questions. METRICS-code-global.md + eval/code-global-qwen.json.

Methodology in ADR-G4. Plan in docs/GLYPH_PLAN.md.

Prerequisites

• Python 3.11+
• ANTHROPIC_API_KEY — only for document extraction by LLM (DocumentExtractor, Claude Haiku 4.5). Code extraction (tree-sitter) and retrieval do not require a key.
• To run the benchmark: a key for any OpenAI-compatible endpoint for judge/generation (NVIDIA NIM, Groq, or Gemini) — see Environment variables.
• The default graph backend is NetworkX (in-process, no server).

Local setup

git clone https://github.com/sammyjdev/glyph-kg.git
cd glyph-kg
pip install -e ".[dev]"
pytest

Architecture

              Extractor port
       +------------+------------+
       |                         |
DocumentExtractor          CodeExtractor
(LLM, probabilistic)       (tree-sitter, deterministic)
       |                         |
       +--------> Graph <--------+
                (Node/Edge)
                    |
              GraphStore port
            +-------+-------+
            |               |
      NetworkX          Neo4j
      (default)        (adapter)
            |
     graph-aware retrieval
            |
   GNOMON benchmark vs vector baseline

Details in docs/ARCHITECTURE.md. Execution plan in docs/GLYPH_PLAN.md.

Environment variables

GLYPH does not define its own env vars — it reads those from the SDKs/endpoints it uses, and the benchmark selects the provider by flag (--base-url/--api-key-env), not by a fixed variable.

Graph backend selection is in code (NetworkXStore default; optional Neo4j adapter), not by env var.

Running the tests

pytest                      # full suite
pytest --cov=glyph          # with coverage
pytest tests/architecture   # architecture invariants

Usage

# 1. Build a code graph (deterministic, no LLM)
from glyph.extract.code import CodeExtractor
from glyph.store import NetworkXStore

nodes, edges = CodeExtractor().extract("path/to/src")     # Python + Java
store = NetworkXStore()
store.upsert_nodes(nodes)
store.upsert_edges(edges)
store.save("out/code.json")

# 2. Serve graph-aware context — the product boundary (ADR-G6), satisfying the Retriever port
from glyph.integration import GraphContextSource
from glyph.embed.sentence_transformers_embedder import SentenceTransformerEmbedder

source = GraphContextSource.from_graph_file("out/code.json", SentenceTransformerEmbedder())
pack = source.retrieve("how is retry handled?", token_budget=1000)   # -> ContextPack
for segment in pack.segments:
    print(segment.score, segment.source, segment.text)

For document graphs, replace CodeExtractor with glyph.extract.document.extractor.DocumentExtractor (requires ANTHROPIC_API_KEY). The global community axis (ADR-G7) lives in glyph.retrieval.community.

Design decisions

ADRs in docs/decisions/:

• ADR-G1: Extractor port + backend selection (NetworkX default, Neo4j adapter).
• ADR-G2: document extraction schema (D&D entities/relations).
• ADR-G3: fair vector baseline + unified output contract.
• ADR-G4: eval methodology (query set, OSS reference-free judge, bootstrap CI, cost).
• ADR-G5: symbol resolution in code extractor (by unique name, declared limitation).
• ADR-G6: GraphContextSource is the product boundary (satisfies the Retriever port).
• ADR-G7: global community axis (Louvain seeded; isolation by pruned traversal projection).

Results and reproduction: METRICS.md (real run, n=25). Technical article (validation-first) in docs/article.md; portfolio claims in docs/portfolio.md.

Relationship to AXON

AXON consumes GLYPH as a dependency: AXON's GraphContextSource (ADR-102 in the AXON repo) delegates to glyph.integration.GraphContextSource from this lib. Code-graph and document-graph become two uses of the same core. Contract and boundary in docs/axon-integration.md.

Contributing

See CONTRIBUTING.md. TDD per sub-task, ADR per architectural decision, invariants checked in CI.

License

Apache-2.0. See LICENSE.