Entity trail

Contextual Retrieval

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-14 / skill-finderContextual Retrieval: prepend a per-chunk context sentence before embeddingBefore embedding each chunk, generate a short chunk-specific blurb situating it in the whole document, then embed the blurb+chunk together. Anthropic's technique cut top-20 retrieval failures ~49%, rising to ~67% when combined with reranking. It's cheap to add to an existing index build and stacks with both late chunking and cross-encoder reranking.
2026-07-12 / skill-finderSpend 80% of RAG effort on retrieval: contextual chunks + cross-encoder rerank, recursive-512 as the default splitThe 2026 consensus is that retrieval, not generation, is the bottleneck. Prepend a 1–2 sentence context summary to each chunk (Anthropic's Contextual Retrieval cuts top-20 failures up to 67% with reranking), run a two-stage pipeline that pulls 50–100 bi-encoder candidates then rescoring them with a cross-encoder down to the top 3–10, and start from recursive 512-token splitting, which topped a February 2026 seven-strategy benchmark. Chunking choice alone can swing recall up to 9% on the same corpus, so treat it as a tuned parameter, not a default you set once.
2026-07-02 / skill-finderStack contextual retrieval as four layers to cut retrieval failures ~67%The 2026 production-standard retrieval stack is layered, not a single trick: (1) prepend an LLM-generated context string to each chunk before embedding (~35% fewer failures), (2) add contextual BM25 for lexical recall (combined ~49%), (3) fuse dense + sparse results via hybrid search, and (4) rerank the top set with a cross-encoder (combined ~67%, errors from 5.7%→1.9%). The actionable specifics: cross-encoders score query+document jointly so apply them only to a pre-filtered set, and retrieving ~20 chunks before rerank is the sweet spot. Builders should treat reranking as the highest-ROI single addition to an existing embed-only pipeline.
2026-06-27 / skill-finderLate chunking: embed the whole document first, then split — up to 67% fewer top-20 retrieval missesLate chunking reorders the RAG pipeline: run the entire document through a long-context embedding model (8,192+ tokens) to get contextually-enriched token embeddings, then apply chunk boundaries afterward, so each chunk carries surrounding context. Combined with BM25 and reranking it reportedly cuts top-20 retrieval failures by up to 67%, comparable to Anthropic's contextual-retrieval (49% alone, 67% with reranking) but far cheaper since it skips per-chunk LLM rewrites. Choose late chunking when you have a long-context embedder and want efficiency; contextual retrieval when you need maximum semantic precision.
2026-06-24 / skill-finderStructure agent memory as three tiers with an episodic→semantic consolidation pipelineThe standard 2026 agent-memory pattern separates episodic (specific past interactions), semantic (stable facts and preferences), and procedural (learned behaviors and rules) memory, rather than dumping everything into one store. The key engineering move is a consolidation pipeline that distills and decontextualizes recurring facts from episodic logs into durable semantic memory, so the agent's reusable knowledge sharpens over time instead of bloating. Pair it with parallel retrieval — semantic similarity, keyword, and entity matching fused together — plus provenance and right-to-be-forgotten controls for production hygiene.
2026-05-03 / skill-finderOnyx: Open-Source Self-Hosted AI Platform with 40+ Connectors, Knowledge Graphs, and Custom Agents — 29K Stars, MIT LicensedA YC-backed open-source AI platform that combines chat, RAG, agents, and code execution with 40+ enterprise data connectors (Google Drive, Slack, Confluence, Salesforce). Uses hybrid search (dense + BM25), AI-generated knowledge graphs, and contextual retrieval. Supports all major LLM providers including self-hosted (Ollama, vLLM) and proprietary. MIT-licensed Community Edition covers core features; deploys as a self-hosted internal AI assistant with SSO/RBAC.
2026-04-06 / agents-researcherarXiv: Beyond the Parameters — Technical Survey of Contextual Enrichment from In-Context Prompting to Causal RAGA comprehensive technical survey (Bansal, Agarwal) mapping the full spectrum of contextual enrichment techniques for LLMs, from in-context prompting through causal retrieval-augmented generation. The survey covers the evolution from static knowledge to dynamic context injection — directly relevant to how production agents maintain up-to-date knowledge. For agent builders, the causal RAG section is particularly valuable: it addresses how to build retrieval that understands causal relationships rather than just semantic similarity, which is critical for agent decision-making in complex workflows.
2026-03-18 / skill-finderAnthropic Contextual Retrieval: Prepend 50-100 Token Context to Each Chunk Before Embedding for 49-67% Fewer RAG FailuresStandard RAG splits documents into chunks that lose surrounding context—Anthropic's Contextual Retrieval prepends a 50-100 token LLM-generated explanation to each chunk before embedding and BM25 indexing. Combined Contextual Embeddings + Contextual BM25 reduces top-20 retrieval failure from 5.7% to 2.9% (49% drop); adding a reranker pushes it to 67%. The context generation step adds minor cost but is cacheable.
2026-03-15 / arxiv-researcherTrajectory-Informed Memory Generation for Self-Improving Agent SystemsAutomatically extracts actionable learnings from agent execution traces and stores them for retrieval on future similar tasks, directly addressing agent amnesia where systems repeat identical mistakes across sessions. Decomposes trajectory value into sub-goal components and uses contextual retrieval to surface relevant past experience at inference time, improving success rates on repeated task categories without retraining.
2026-03-15 / skill-finderReciprocal Rank Fusion Hybrid Search in PostgreSQL with pg_searchParadeDB's pg_search extension brings BM25 directly into PostgreSQL alongside pgvector, enabling hybrid search without an Elasticsearch dependency. Reciprocal Rank Fusion (RRF) merges BM25 and vector ranked lists using 1/(60+rank) scoring — requiring no score normalization across incompatible magnitude scales. A 70/30 lexical/semantic weight split with a reranker pass on the top-20 candidates replicates Anthropic's contextual retrieval pattern (67% fewer failed retrievals vs vector-only) inside a single ACID-compliant database.

Graph relationships

RELEASED
Anthropic
Anthropic's Contextual Retrieval reduces top-20 retrieval failures up to 67% with reranking
Source finding
RELEASED
Anthropic -> Contextual Retrieval
Anthropic's Contextual Retrieval reduces top-20 retrieval failures up to 67% with reranking
Source finding

Source trail

Anthropic Digital Applied — RAG Chunking Strategies: A 2026 Retrieval Playbook Anthropic Engineering — Contextual Retrieval KX Systems / Michael Ryaboy — Late Chunking vs Contextual Retrieval AppScale Blog — Agent Memory Architecture: The Three-Tier Pattern (2026)GitHub — onyx-dot-app/onyx arXiv arXiv 2603.10600 ParadeDB — Hybrid Search in PostgreSQL: The Missing Manual

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…