Granveo field notes

Notes on building useful context.

Research-backed essays about knowledge graphs, AI memory, retrieval, and the work of turning scattered evidence into decisions.

A source card connected to a traceable working claimTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question

Latest writing

One careful piece at a time.

Several source records linked through evidence and review decisions to a canonical entity, while unresolved and different records stay separateSOURCESConnected contextrelationships stay visibleSYNTHESIS
Knowledge graphs11 min read

How to Resolve Entities Without Corrupting a Knowledge Graph

Entity resolution is not a cleanup step. It is a continuing decision about which records describe the same real-world thing, backed by evidence and designed to be undone.

Read the field note
Two insurance policy packages aligned into clause-level differences with citations back to schedules, wordings, and endorsementsTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Insurance knowledge workflow10 min read

How to Compare Insurance Policy Wordings Reliably with AI

Reliable policy comparison is not a document summarization task. It aligns the whole policy package clause by clause, makes additions and omissions explicit, and gives a reviewer the exact evidence behind every reported difference.

Read the field note
A source change moving through detection, versioning, parsing, indexing, retrieval, and freshness validationA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Retrieval and context10 min read

How to Keep a RAG Knowledge Base Fresh as Sources Change

A fresh RAG knowledge base is a change-processing system. It must detect revisions, rebuild every dependent artifact, preserve history, and prove that deleted or superseded evidence can no longer answer current questions.

Read the field note
A broad retrieval set narrowed through relevance, diversity, and evidence checks into a compact source-cited contextTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How to Rerank RAG Results Without Hiding Retrieval Failures

Reranking should turn a broad candidate set into a compact, useful evidence set. It cannot recover missing evidence, repair bad permissions, or make a weak source authoritative.

Read the field note
A question traced through page images, document text, charts, diagrams, and layout regions to a source-cited answerTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How to Build Reliable Multimodal RAG for Visual Documents

Visual documents are not text files with decorative pictures. Reliable multimodal RAG preserves each page, indexes several representations, retrieves evidence across modalities, and cites the exact page or region behind an answer.

Read the field note
An email question traced through messages, quoted replies, attachments, and decisions to a chronological source-backed answerTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How to Build RAG for Email and Conversation Threads

Email RAG should reconstruct conversations, not turn every message into an unrelated text chunk. The useful unit is a source-linked thread of messages, attachments, people, decisions, and changing state.

Read the field note
A multilingual question branching into original-language, translated, and cross-lingual retrieval paths that reconnect at cited source evidenceSOURCESConnected contextrelationships stay visibleSYNTHESIS
Retrieval and context10 min read

How to Build Multilingual RAG Without Losing Meaning

Multilingual RAG is not ordinary RAG with a translation step. It needs explicit language routes, terminology controls, cross-lingual retrieval, and citations that lead back to the original evidence.

Read the field note
A spreadsheet question routed through schema retrieval, row filtering, calculation, and cell-level citation before producing an answerA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Retrieval and context10 min read

How to Build RAG for Spreadsheets and Tabular Data

Spreadsheet RAG should not treat a workbook as a bag of text. It needs to preserve table structure, find the right rows and columns, execute calculations, and return an auditable cell-level evidence trail.

Read the field note
Two sources making scoped claims about the same relationship, with derivation, version, and review paths preserved in a knowledge graphTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Knowledge graphs10 min read

How to Build a Knowledge Graph That Preserves Evidence

A trustworthy knowledge graph does not turn every extracted relationship into a fact. It records who asserted each claim, what evidence supports it, how it was derived, and where its scope ends.

Read the field note
A complex question split into connected retrieval steps that join evidence from several sources into one traceable answerSOURCESConnected contextrelationships stay visibleSYNTHESIS
Retrieval and context10 min read

How to Build RAG for Questions That Need Multiple Sources

Multi-source questions fail when retrieval treats every passage as an independent match. Reliable multi-hop RAG makes the evidence chain explicit, retrieves each dependency, and shows where the chain breaks.

Read the field note
A RAG question passing through retrieval, evidence sufficiency, correction, and claim support checks before either a grounded answer or a useful abstentionTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How to Make RAG Systems Abstain When Evidence Is Weak

A reliable RAG system needs more than a refusal prompt. It must recognize when retrieved evidence cannot support the requested claim, try bounded recovery, and return a useful abstention without hiding what failed.

Read the field note
A source document moving through format detection, layout-aware extraction, structural validation, versioning, and traceable RAG retrieval unitsA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Retrieval and context10 min read

How to Build a Reliable RAG Document Ingestion Pipeline

Reliable RAG starts before embeddings. A sound ingestion pipeline keeps the original file, extracts a structured document, tests difficult content, and preserves a trace from every retrieval unit back to its source.

Read the field note
Source documents carrying identity, date, version, scope, and provenance metadata through filters into a traceable RAG answerTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How to Design Metadata That Makes RAG Retrieval Better

Useful RAG metadata does more than label a file. It helps the system decide which evidence is eligible, relevant, current, and traceable to its source.

Read the field note
A question branching toward a bounded long-context document set or a retrieval path, then reconnecting both routes to cited source evidenceSOURCESConnected contextrelationships stay visibleSYNTHESIS
Retrieval and context10 min read

Long Context or RAG: How to Choose the Right Approach

A larger context window does not automatically replace retrieval. The better design depends on how much evidence exists, how it changes, and how answers must be reviewed.

Read the field note
Two disagreeing source paths carrying dates and provenance into a reviewed, qualified RAG answerTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How RAG Systems Should Handle Conflicting Evidence

Conflicting evidence is not retrieval noise to average away. A reliable RAG system identifies the disputed claim, compares source context, and keeps unresolved disagreement visible.

Read the field note
An authenticated user passing through an authorization gate before permitted sources enter a traced RAG contextTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and context10 min read

How to Build Permission-Aware RAG Without Data Leaks

A secure RAG system does not retrieve broadly and ask the model to stay quiet. It limits the searchable evidence before ranking, then verifies access again before context assembly.

Read the field note
Keyword and vector retrieval paths merging into a ranked, filtered evidence set for a source-grounded answerA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Retrieval and context10 min read

How Hybrid Search Improves Retrieval for RAG Systems

Hybrid search gives exact terms and semantic similarity separate ways to earn a place in the context. The hard part is combining them without hiding retrieval failures.

Read the field note
A document flowing through structural parsing, retrieval units, context assembly, and a source-grounded answerA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Retrieval and context10 min read

How to Choose the Right Chunking Strategy for RAG Systems

Good chunking is not a universal token count. It is a decision about which unit should be found, read, cited, and reconnected to its source.

Read the field note
A RAG evaluation trace separating a user question, retrieved evidence, grounded claims, and a reviewed answerTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Retrieval and evaluation10 min read

How to Evaluate a RAG System Without Fooling Yourself

A useful RAG evaluation does not collapse retrieval and generation into one score. It shows which part failed, on which questions, and why.

Read the field note
A long-term memory workflow connecting observations, typed records, retrieval, action, review, and forgettingA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Context engineering10 min read

How to Build Reliable Long-Term Memory for AI Agents

Reliable agent memory is not a transcript archive or a larger prompt. It is a controlled system for deciding what to preserve, what to retrieve, what to update, and what to forget.

Read the field note
An insurance journey connecting product explanation, proposal, policy, servicing, claim, and decision evidenceA DURABLE RESEARCH LOOP1Collectfiles + notes2Connectclaims + links3Questiongaps + tension4Createbrief + draftnew questions return to the map
Insurance knowledge10 min read

Why Insurance Penetration and Trust Stay Low in India

India's insurance gap is not one problem with one fix. Affordability, product salience, sales incentives, fragmented policy knowledge, and claims experience reinforce one another.

Read the field note
A research question connected to evidence cards, checked claims, and a source-grounded draftTRACE A CLAIMSOURCE 03WORKING CLAIMContext is more usefulwhen its origin survives.3 sources · 2 relationships · 1 open question
Research workflows10 min read

How to Build a Source-Grounded AI Research Workflow

A source-grounded workflow does not ask AI to write from a pile of documents. It separates evidence capture, claim building, drafting, and verification.

Read the field note
Research sources flowing into a connected context graph and a source-grounded synthesisSOURCESConnected contextrelationships stay visibleSYNTHESIS
Context engineering11 min read

What Is an AI Knowledge Graph? Connected Context Explained

A knowledge graph does more than store information. It keeps the relationships and source trail that make information useful when the question changes.

Read the field note