Connected policy knowledge and health¶

Status: Draft | Last reviewed: 2026-10-09

This proposal covers topic modeling, cross-repository terminology, AI-assisted graph construction, and the effectiveness of combined indexes. Use it when the answer requires definitions, conditions, exceptions, and evidence from several documents rather than a single matching passage.

The CDC and internal-policy scenario¶

Use selected live CDC pages and an organization's private policy repository as the main acceptance scenario. CDC's material includes both infection-control guidance and pages describing legal authorities and regulations (C1, C2, C3). The source connector must preserve that distinction rather than classify every CDC page as a regulation.

An internal policy may use a local name for a concept described differently by CDC, refer to another internal procedure for exceptions, and quote an older CDC version. Useful output must resolve those connections and make the uncertainty visible.

For a concrete terminology fixture, use CDC's Standard Precautions and Transmission-Based Precautions, which CDC describes as distinct tiers used together (C1, C2). A fictional internal document can call a procedure “routine care precautions.” Its relation to either CDC concept must be established from the procedure's definition and scope; the local name alone proves nothing. The fictional internal wording is a test fixture, not an observed organization policy or an asserted clinical equivalence.

The author should be able to ask for:

A crosswalk between external concepts and internal terminology.
A supported comparison of external guidance and internal policy.
All applicable definitions, conditions, exceptions, and unresolved references for a scenario.
Internal policies potentially affected by an external section changing.
Topics whose evidence is missing, stale, conflicting, or weakly retrievable.

Capture requested dates, jurisdiction, setting, audience, and policy version when relevant. Ask for missing applicability context or report separate possibilities. Do not invent a universal precedence order between external and internal sources.

Four connected representations¶

Representation	Purpose
Source/structure store	Original documents, revisions, sections, tables, citations
Terminology and topic catalogue	Concepts, labels, mappings, definitions, topic membership
Evidence graph	Supported assertions, applicability, definitions, exceptions, cross-references
Vector/keyword index	Find useful entry points and evidence when language differs

These are logical, versioned views of one collection, not four required database services. The graph does not replace documents, and similarity does not establish equivalence.

Use RAGFlow's native graph capabilities first, alongside its vector/hybrid index. Its GraphRAG already provides extraction and entity deduplication R1; current Knowledge Compilation adds configurable graph types and extraction rules R2. TrueServe should supply the authoring, review, evidence, evaluation, and publication experience around that engine.

The native capability proof tests the selected version against the policy contract below. Add a sidecar, native extension, or separate graph adapter only for a demonstrated gap. Source lineage and review history remain portable without requiring a full duplicate graph.

The authored processing graph remains acyclic. The domain knowledge graph may contain cycles, reciprocal references, and several paths to the same concept. Its query executor needs visited-node handling and bounded traversal.

Topic modeling¶

Support both an authored taxonomy and discovered topic candidates.

Create/import a hierarchy with stable topic IDs, descriptions, and examples.
Assign sections/clauses to one or more topics with evidence.
Cluster unmatched or broad material to suggest new topics or subdivisions.
Let an author merge, split, rename, or reject suggested topics.
Track topic membership, source coverage, and retrieval results by release.

An authored topic can remain stable while its label or clustering model changes. Store discovered clusters separately until they are mapped to approved topics. Allow multi-topic membership: an exception may concern several policy areas.

Evaluate native RAGFlow compilation for topic organization and hierarchical summaries first; its current artifacts include Tree, Mind Map, and Wiki R3. BERTopic offers hierarchical/time-based analysis (T1, T2); Microsoft GraphRAG offers community summaries and local entity-centered retrieval G1. Treat these as alternatives to qualify when native results leave a measured gap.

Topic/community summaries help navigation, discovery, and broad synthesis. An applicability or exception claim still needs the underlying source evidence. Do not turn a topic assignment into an access rule or an automatic exclusion from retrieval.

Compare topic changes only with their model/version context. A new clustering model can rename or rearrange topics without any source-policy change. Show that separately from content drift.

Terminology alignment¶

Give every concept a stable ID, preferred labels, language-tagged aliases, definition, source scheme, examples, and scope. Preserve the source's original wording in evidence and citations.

Use an explicit mapping vocabulary inspired by SKOS, which distinguishes exact, close, hierarchical, and related mappings K1:

Mapping	Retrieval behavior
Exact	Approved equivalence within the declared scheme/scope; eligible for alias expansion
Close	Useful candidate expansion, with the distinction retained
Broader/narrower	Directional expansion that preserves which meaning is more specific
Related	Explore a connection without substituting the term
Different/ambiguous	Prevent a known false match or request disambiguation

Every mapping records source terms/concepts, language, definition evidence, scope, mapping kind, proposed confidence if available, reviewer status, and effective/review dates. Importing a vocabulary does not authorize every cross-vocabulary equivalence.

Generate mapping candidates with lexical normalization, abbreviation lookup, definition embeddings, structural context, and optionally an LLM comparison. Require the comparison to identify supporting definitions and differences. Similarity alone cannot approve an exact mapping.

Reuse native entity resolution where exposed, but verify its scope and merge behavior. An automatic deduplication result is a mapping candidate; it must not override reviewed distinctions between external and internal policy terms.

Do not collapse “patient,” “member,” and “insured person” into one entity merely because they appear in related documents. The roles may overlap in a scenario while retaining different policy meanings. Acronyms and translations also need repository/context scope.

The terminology review UI shows external and internal definitions side by side, suggests mapping type, and previews affected retrieval cases. It supports approval, rejection, merge/split, expiry, and remapping without deleting original source terms.

At query time, keep the original wording plus a trace of selected expansions. An ambiguous term can search several interpretations and label the results, or ask for clarification. Mapping revisions invalidate dependent caches and trigger targeted evaluations; changes to indexed retrieval text create a new index build.

Evidence graph schema¶

Start with a small configurable policy schema. Add domain-specific types through a visual schema editor with field descriptions, validators, and examples.

Node	Example fields
Source document/version	Authority, document kind, publication/effective/review dates, status
Section/clause/table	Structural locator, exact extracted text, parent, source revision
Concept/term	Definition, labels, scope, vocabulary version
Assertion	Subject, predicate, object/value, polarity, modality, conditions, provenance
Topic	Stable identity, description, approved hierarchy
Scenario/applicability scope	Product, setting, population, geography, date constraints

Representative relations are DEFINES, REFERS_TO, APPLIES_TO, CONDITION_OF, EXCEPTION_TO, QUALIFIES, ABOUT, and explicitly supported SUPERSEDES. Candidate CONFLICTS_WITH relationships need evidence from both sides and an unresolved/resolved state.

Represent assertions as addressable artifacts using the native graph schema or a small evidence sidecar where needed. A bare subject-predicate-object triple cannot carry all the conditions, negation, source version, or review state required for policy interpretation.

Retain:

Exact source span/block references for the relation and important field values.
Normative modality, such as requirement, recommendation, permission, or prohibition, without strengthening the source language.
Conjunction/disjunction, thresholds, units, temporal qualifiers, and exceptions.
Document authority and applicability scope, with extracted versus approved values distinguished.
Source-valid dates separately from acquisition/extraction times.
Extraction model/prompt/schema version and candidate/verified/rejected status.

Unknown dates or scope remain unknown. A later web fetch time is not a legal effective date, and a newer document is not automatically a superseding policy. Keep any executable decision rules as separately authored/versioned artifacts; AI extraction must not silently create adjudication logic.

AI construction and review¶

Add first-class preparation nodes:

Identify topics → Extract concepts/definitions → Propose term mappings → Extract assertions/relations → Resolve references → Validate graph → Index graph.

These nodes can branch from structured documents before chunking and rejoin the knowledge bundle through stable source anchors. Graph assertions should survive a different text splitter without inventing new meanings.

Use native extractors/compiler stages behind these steps where supported. If a native job combines several operations, show it as one managed stage with its actual inspection boundaries; do not invent independently editable checkpoints.

Each AI node exposes model, full prompt, input unit, extraction schema, typed field mappings, and its validation/review policy. AI can populate graph fields and propose relationships; identity, tenant access, and publication status stay under system control.

Validate source-span existence, endpoint identity, relation types, field types, conditions, and scope. A matching quote proves that words exist, not that the extracted relation follows from them. Add entailment checks and human review for critical or ambiguous mappings/assertions.

Candidate edges may help locate source evidence, but they must not silently establish equivalence, authority, precedence, or a policy conclusion. Keep that discovery use distinct from validated/reviewed knowledge.

Resolve explicit section references deterministically where possible. Leave an unresolved-reference record when the target is unavailable; do not fill it with a model guess. Treat relation confidence and review status as different properties.

Deduplicate candidate concepts using definition/context evidence. Preserve separate entities until a merge is approved. Keep merge/split history so a false equivalence can be reversed and affected indexes rebuilt.

Retrieval and policy synthesis¶

The Answer canvas expands to a bounded evidence plan:

Identify the requested task, terms, constraints, and missing context.
Resolve terms against the versioned vocabulary without discarding ambiguity.
Run keyword/vector retrieval over authorized content to find entry points.
Expand approved graph relationships: definitions, referenced clauses, conditions, exceptions, and relevant source versions.
Retrieve the actual supporting source passages for the selected graph facts.
Check applicability, conflicting support, missing dependencies, and context coverage; perform a bounded further retrieval pass when configured.
Synthesize a response with a claim/evidence matrix, citations, scope, and unresolved points.

Use RAGFlow's native graph-assisted retrieval first. Distinguish returned graph-derived chunks from an explicit evidence path: enabling graph retrieval alone does not prove that bounded traversal or complete provenance is exposed. Offer controls and traces that the selected native adapter has verified.

Bound hop count, visited nodes, relation types, branch width, model calls, context, and latency. Filtering must apply throughout the path, including intermediate nodes and edge evidence, before any model sees it.

A route to an assertion is not sufficient support for the assertion. Every material answer claim must resolve to authorized source evidence with the relevant qualifiers. Derived graph/community summaries remain labeled.

If a required definition, exception, source version, or applicability fact cannot be resolved, return a partial analysis with that gap. Do not silently convert “required evidence unavailable” into either compliance or noncompliance.

For broad topic synthesis, use topic/community retrieval. For precise policy interpretation, prioritize source clauses and their dependencies. A single retrieval strategy should not be forced onto both tasks.

Health workspace¶

Measure effectiveness with evidence tasks, not only short-answer Q&A. Tasks include policy comparison, scenario interpretation, cross-source synthesis, definition matching, conflict analysis, and change impact. Each task can specify required claims, qualifying/counter-evidence, acceptable scope, and structured output assertions.

Health dimension	Signals	Drill-down/action
Source coverage	Missing/stale documents, incomplete refresh, unavailable references	Affected sources, topics, tasks
Structural fidelity	Lost clauses/tables, unsupported locators, missing qualifiers	Original versus parsed evidence
Topic coverage	Unclassified material, thin source coverage, failing tasks by topic	Topic/source matrix and gap queue
Observed demand	Transcript-derived question groups, reviewed support, unresolved needs by driver/topic	Question occurrence evidence and prioritized knowledge gaps
Terminology	Unmapped terms, rejected/ambiguous equivalences, mapping drift	Definition crosswalk and affected cases
Graph integrity	Orphan/dangling references, unsupported assertions, duplicate concepts	Source-backed neighborhood inspection
Applicability/versioning	Stale assertions, inconsistent scopes, unresolved supersession	Source/version comparison
Retrieval effectiveness	Complete evidence-set recall, path coverage, missed exceptions	Query and traversal trace
Synthesis	Supported-claim coverage, qualifier retention, contradictions, useful abstention	Claim/evidence matrix
Publication consistency	Graph/vector/source generation agreement	Repair/rebuild before activation
Operations	Freshness lag, queue age, resource use, provider failures	Worker/source operations

Graph density, node count, and connectedness are diagnostics, not quality targets. Adding unsupported edges can improve connectivity while making answers worse. Separate measured failures from heuristic warnings and unknown coverage.

Use transcript discovery to measure coverage against actual requests as well as document-generated questions. Show distinct-question and demand-weighted results, source population, sampling, and unreviewed coverage. Call Drivers findings can prioritize investigation; an unresolved contact does not by itself establish a knowledge gap or a causal effect of the index.

Pin a corpus/release scope and evaluation window in every health view. Display the number and origin of evaluated tasks. A topic with no verified tasks is unmeasured, not healthy. Drill-down should lead to a fix and the tasks that must be rerun.

Change impact and publication¶

Keep raw CDC snapshots, canonical URLs, redirects, declared dates, fetched times, and source-type classifications. Follow configured site sections and linked documents; do not treat navigation or a cosmetic page edit as changed policy.

Compare content at section/clause level. Rebuild assertions and mappings that depend on changed passages. Traverse their reverse dependencies to identify potentially affected internal provisions and evidence tasks.

Show the external diff, affected internal passages, mapping assumptions, confidence/review state, and the proposed reason for impact. This is a review queue; it does not edit or publish the organization's policies.

Version topic models, vocabularies/mappings, assertions, vector indexes, and native graph artifacts together in the publication manifest. A graph-enabled query must not combine a new vector index with an unrelated old graph. Verify every native artifact required by the selected query profile before publication.

Compare a hybrid-only baseline with native graph retrieval, then measure the additional effects of reviewed terminology expansion and topic summaries. Retain the configuration, evidence, latency, and cost for each experiment.