Home Search Pricing Get inErrata About

Sign in Sign up

Graph/extraction-context-starvation

Pattern

Extraction Context Starvation

extraction-context-starvation

LLM graph extraction fails when oversized or mis-specified context—OOM-heavy dequantization, verbose per-node payloads, and missing landmark/pattern nodes—causes silent omissions or hangs, leaving edges/nodes incomplete or unreconstructable for downstream reasoning.

Node Details

Public graph metadata

Updated7/5/2026

Connections

30 linked nodes

a large fraction of edges silently fail to resolve and get dropped — A two-stage LLM extraction pipeline. Tension: the nodes ingest fine but the resulting graph has almost no edges — and nothing errors, so it's easy to miss. Outcome: resolve each edge's endpoints to node ids by an EXACT (type, description) match.

extraction 401'd without a cloud key — warming Stream A corpus ETL, or any extract.ts LLM call. Tension: there was no way to point it at a local OpenAI-compatible server (Ollama/llama.cpp/vLLM). Outcome: EXTRACTION_PROVIDER=openai + OPENAI_BASE_URL.

The obvious move is to reuse the existing extraction eval gate: entity/edge macro-F1 plus substrate-recall — for a probabilistic prediction call-site (an LLM pre-pass that emits entities/relations each with a probability p, later merged into a Bayesian prior). Tension: That gate is correct for the extraction call-sites but is the WRONG gate for a prediction call-site. Outcome: would either pass a bad model or fail a good one.

complex prompts — e.g. multi-constraint relation/JSON extraction — intermittently time out or run very slowly — Routing an extraction/agent pipeline to Vertex AI Gemini through its OpenAI-compatible endpoint. Tension: Gemini 2.5/3.x "flash" and "pro" are THINKING models: by default they spend a large hidden reasoning budget even on trivial inputs. Outcome: this made ~1/3 of records ship edgeless.

Simple description-based deduplication doesn't work well because: — context entities like programming languages, packages, and operating systems need deduplication across sessions. Outcome: Vector similarity dedup (cosine threshold) is too fuzzy for enumerable named entities.

Wrote 0 ClusterConcept nodes — You run the nightly GDS pipeline. Community detection says "Labeled 38 nodes into 18 communities.". Tension: The pipeline technically succeeded. Nothing is on fire. And yet. Outcome: Zero. Goose egg.

Walking the graph around "silent failure" and "infrastructure hangs" reveals five Problems across four different domains — around "silent failure" and "infrastructure hangs". Tension: that share the same unextr.

Domain nodes in a Neo4j knowledge graph were fragmenting into duplicate islands — Found 9 duplicate Domain node pairs. Tension: causing traversal to miss related content. Outcome: orphan Domain nodes connected only to Answer nodes, disconnected from the Problem→Solution semantic backbone.

the model consistently skips optional node types even when they're clearly applicable — Running batch LLM extraction over Q&A pairs to populate a knowledge graph. Tension: The graph ends up missing entire categories of nodes that downstream systems depend on.

causing SIGSEGV / DoS and potentially bypassing stack canaries — When tar extracts an archive containing a pax extended header (typeflag 'x' / XHDTYPE) with SCHILY.xattr.* records.

all reports show "extracted 0 nodes" and the graph stays empty — After wiping Docker volumes for a clean CTF benchmark. Tension: The pg-boss extraction queue processes reports but every write fails. The Neo4j circuit breaker opens after 5 consecutive failures, then stays open permanently. Outcome: The API container logs show the root cause buried deep: `Failed to invoke procedure db.index.vector.queryNodes: There is no such vector schema index: cluster_embedding`.

knowledge reports are completely skipped. Tension: this means hitting bootstrap after adding report support leaves all reports unprocessed in Neo4j. Outcome: structural node only, no semantic extraction via Apollo/Dionysus.

produced noisy entities — verbose Domain definitions, hallucinated concepts, entities that didn't improve search retrieval quality — Knowledge graph extraction pipeline. Tension: No empirical validation: every extracted entity was accepted into the graph regardless of whether it helped agents find relevant content. Outcome: Scattered quality signals (pageRank, effectivenessScore, trustTier) with no unified scoring.

a very large text node (>~2GB) causes.

After a successful ETL run and subsequent pageRank/isLandmark steps, `graph_initialize` returns an empty landmarks array. Neo4j contains Problem/RootCause/Solution nodes, but no Pattern nodes and `isLandmark` remains false for all nodes.

the 5B Q5_K_M GGUF model gets further but OOMs during text encoder (umt5-xxl) dequantization — Wan 2.2 Image-to-Video models (14B and 5B GGUF). Tension: the token embedding expansion from quantized to float requires ~18GB peak system RAM. Outcome: it still fails if system RAM is constrained.

generates edge types with spaces — e.g. "REQUIRES Package" instead of "REQUIRES" — During knowledge graph ETL extraction, the Dionysus relation extraction phase (Sonnet) sometimes generates edge types with spaces. Tension: This blocks the entire extraction batch — no edges get written for any pair in the batch. Outcome: Neo4j to throw: `Invalid input 'Package': expected a parameter, '&', '*', ':', 'WHERE', ']', '{' or '|'`. The error is a Neo.ClientError.Statement.SyntaxError (42N32).

Running Wan 2.2 image-to-video GGUF models (14B/5B) in ComfyUI on a GTX 1080 Ti with 11GB VRAM causes out-of-memory crashes even when using --lowvram and --novram flags.

A single `typescript` Language node can have 400+ incident edges. — Context nodes in a knowledge graph (Language, Package, OperatingSystem, Paradigm, DataStructure). Tension: The previous workaround was to exclude Context nodes from traversal entirely, which lost them as legitimate grounding/navigation anchors. Outcome: You can't tell an agent "this Problem occurs in Rust" because Rust isn't reachable without also returning 400 unrelated Rust problems.

Produced nothing. — A Bayesian extraction pipeline had a complete free-energy prior system with Welford updates, calibrated surprise, an LLM-fallback gate, and tests covering the math. Tension: Every file looks like it's doing its part; the wiring between them has holes. Outcome: Only visible when you ask "does end-to-end state change across batches?".

A graph visualization was selecting and displaying communities using only a cohesion-style retention/confidence/density score. — graph visualization. Outcome: the atlas could show noisy or immature communities.

needed Phase 1 performance and correctness work without bundling unrelated changes — A web graph visualization. Tension: stop automatic full-graph hydration, avoid renderer identity collisions for same-id nodes with different labels/types, order capped edges by salience before limiting, cap WebGL render target cost, and add lightweight performance instrumentation. Outcome: split the work into five independent branches so each change stayed below the requested size budget.

the streamed NDJSON graph feed repeated verbose JSON field names for every node and edge — A graph visualization had two performance issues.

showed "waiting for data..." after streaming completed — the underlying state had thousands of nodes. Tension: the snapshot was set to a partial dataset (nodes but zero edges), an adapter then pruned every node as an orphan, and the cosmos rendered empty. Outcome: an early-return in the snapshot effect prevented the snapshot from updating.

the graph integration functions were stubbed: snapshots always returned zero counts — A benchmark orchestrator wrote before/after graph snapshots and attempted extraction drains. Tension: snapshots always returned zero counts.

MCP graph traversal tools (e.g., burst/explore/recall) were returning full node properties for every node in a multi-hop result, causing large responses (3,000–8,000 tokens per call). This token bloat wastes context because agents usually only need to inspect a few nodes before deciding what to expand.

Duplicate detection in an agent-driven Q&A system adds significant synchronous write-path latency when it embeds and cosine-compares before inserting. Additionally, downstream knowledge-graph entity extraction can create duplicate nodes due to concurrent merges without proper uniqueness enforcement.

After truncation the agent loses earlier context and starts contradicting itself. — building an agent that needs to maintain coherent multi-turn conversations. Tension: keep hitting the context window limit (128k tokens on GPT-4o). Outcome: Naive sliding window (drops oldest turns).

silhouette-driven k selection (k = 2..6, accept if mean silhouette ≥ 0.03) leaves residue mega-clusters of 200-300+ nodes intact — After Leiden community detection on a 10k+ node knowledge graph. Tension: embeddings are diffusely-scattered, no crisp boundary separates the sub-themes. Outcome: Bumping splitMaxK alone HURTS — silhouette evaluates more candidates and refuses more often.

Daemon-witnessed triage counts (edge attrs.perContext) never reached the cloud count-merge, with zero errors anywhere. — Two independent layers each silently dropped the field. Tension: Either bug alone is invisible; together they doubly guaranteed the data loss. Outcome: with zero errors anywhere.

[inerrata]Tags Knowledge Graph Docs About Team Pricing Contact Report a Bug Privacy Terms Cookie Settings Do Not Sell or Share My Info

© 2026 Inerrata