Build

Knowledge Base β€” Strategy & Policy

15 min read

🧠 Indexa Knowledge Base β€” Strategy & Policy

What this document is. The constitution for Indexa’s content knowledge base. It defines how we mirror the 82 published articles into Obsidian, how we decompose them into a reusable peptide knowledge graph, how that graph is governed and ingested into Supabase, and how downstream channels (TikTok, Threads) borrow from it without ever polluting it. Read this before building anything. Once approved, every scraping run, note, and content piece must comply with it.

1. North Star

Build one canonical, interlinked body of peptide knowledge that is true, traceable, and reusable β€” so that:

  • Customers and researchers can learn from it (surfaced via the site / Supabase).
  • Every social asset (TikTok, Threads, captions, emails) is a derivative of it, never an original source.
  • The knowledge compounds: each new paper, anecdote, or protocol enriches existing nodes instead of spawning duplicates.

The governing principle: The knowledge base states what is true. Channels decide how to say it. Truth lives in one place; tone, voice, and platform packaging are applied on the way out, never baked in.

2. Three-Layer Architecture

LAYER 1 β€” SOURCE MIRROR        LAYER 2 β€” KNOWLEDGE GRAPH       LAYER 3 β€” CHANNEL OUTPUT
(faithful copy of the 82       (atomic, deduplicated,          (TikTok / Threads / email β€”
 published articles)            evidence-tagged truth)          tone-manipulated derivatives)

  Article notes  ───extract──▢  Peptide / Mechanism / Paper  ──manipulate──▢  Content posts
  (1:1 rebuild)                 Protocol / Concept notes                      (alias brands, hooks)

         └──────────── Obsidian = source of truth β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜
                                  β”‚ ingest
                                  β–Ό
                          Supabase (read model)
  • Layer 1 β€” Source Mirror. A faithful 1:1 rebuild of each published article. This is the provenance anchor β€” proof of what we already say publicly, and the raw material for extraction. We do not editorialise here.
  • Layer 2 β€” Knowledge Graph. Atomic notes (one peptide, one mechanism, one paper, one protocol = one note) that articles link into. This is the reusable, dedupe’d truth that gets enriched over time and ingested to Supabase.
  • Layer 3 β€” Channel Output. Lives in the existing Content/ pipeline, not in the knowledge base. Pulls facts from Layer 2, then applies channel tone/policy. Covered by Content β€” Playbook.

Hard rule: Layers 1 and 2 are the source of truth. Layer 3 never writes back into them. A TikTok hook is not allowed to change what a peptide note says.

3. The Hybrid Note Model

Every piece of knowledge is one of six note types. The first two are Layer 1; the rest are Layer 2.

TypeTagOne note =Purpose
Article#kb-articleone published blog post (1:1 mirror)Provenance + the assembled, customer-facing narrative
Resource#kb-resourceone site resource/guide (dosage, reconstitution, storage, dictionary)Mirror of evergreen tool/guide pages
Peptide#kb-peptideone compound (e.g. BPC-157)The canonical entry for a molecule β€” the hub everything links to
Mechanism#kb-mechanismone biological pathway/action (e.g. angiogenesis, GH secretagogue)Cross-functional concept reused across peptides
Paper#kb-paperone study / citation (PubMed/PMC)Evidence node β€” the acute breakdown of a single paper
Concept#kb-conceptone reusable idea (e.g. reconstitution, COA, half-life, stacking)Glossary-grade building blocks
Transcript#kb-transcriptone external source (podcast/talk/interview transcript)Layer-1 anecdote source β€” provenance: anecdote; its claims thread into peptide notes’ Anecdotes/Protocols sections
Source#kb-sourceone reputable external reference (journal/conference/oncology network)Higher-tier external evidence (provenance: literature) β€” e.g. ASCO/JAMA findings; cited like transcripts but weighted above anecdote
Supplement#kb-supplementone non-peptide nutritional adjunct (CoQ10, NAD, methylene blue, glycine)Adjunct entity β€” dose/route/timing/intake/mechanism; linked from stacks
Compound#kb-compoundone non-peptide therapeutic (MK-677, orforglipron, enclomiphene, clomiphene, cardarine)Non-peptide drugs (SERMs, small-molecule GH/GLP-1 agonists, etc.) that act on the same mechanisms as peptides but aren’t peptides β€” kept honest rather than mislabeled kb-peptide
Behavior#kb-behaviorone lifestyle/behavioral intervention (sauna, cold plunge, Zone 2, HIIT, fasting, sleep, sunlight)Behavioral layer β€” the non-pharmacological inputs that compose into stacks alongside peptides/supplements
Stack#kb-stackone combination protocol (peptides + supplements + behaviors)Layer 3 (applied) β€” synergy rationale, timing, intake, cycling, cautions, monitoring
Method#kb-methodone cross-cutting methodology/reference (e.g. animal→human dose conversion)Reference playbooks for working with the data — the convention for turning animal-study doses into traceable human-equivalent breakdowns

Why hybrid (vs. 1:1-only or fully-atomic): the article mirror preserves SEO structure and provenance; the atomic layer is what makes the base a graph β€” reusable, enrichable, and clean to ingest. Articles become thin assemblies that link to fat, canonical concept notes. One fact (e.g. BPC-157’s angiogenesis mechanism) is written once in the Peptide/Mechanism note and referenced by every article that touches it.

Anti-duplication law: if a fact belongs to a peptide, mechanism, paper, or concept, it is written in that atomic note and linked β€” never re-typed inside an article. Articles narrate and contextualise; atomic notes hold the truth.

4. Folder Structure

New home: 01 - Distribution/Knowledge Base/. This is the β€œcompile all content pillars” folder.

01 - Distribution/
└── Knowledge Base/
    β”œβ”€β”€ Knowledge Base β€” Strategy & Policy.md   ← this file
    β”œβ”€β”€ _Templates/
    β”‚   β”œβ”€β”€ _Template β€” Article.md
    β”‚   β”œβ”€β”€ _Template β€” Peptide.md
    β”‚   β”œβ”€β”€ _Template β€” Mechanism.md
    β”‚   β”œβ”€β”€ _Template β€” Paper.md
    β”‚   └── _Template β€” Concept.md
    β”œβ”€β”€ Articles/            ← Layer 1: 86 mirrored posts
    β”œβ”€β”€ Resources/           ← Layer 1: dosage, reconstitution, storage, dictionary
    β”œβ”€β”€ Transcripts/         ← Layer 1: external anecdote sources (podcasts/talks)
    β”œβ”€β”€ Sources/             ← Layer 1: reputable external references (journals, conferences, oncology networks)
    β”œβ”€β”€ Peptides/            ← Layer 2: one note per compound
    β”œβ”€β”€ Mechanisms/          ← Layer 2: pathways / actions
    β”œβ”€β”€ Papers/              ← Layer 2: cited studies
    β”œβ”€β”€ Concepts/            ← Layer 2: glossary building blocks
    β”œβ”€β”€ Supplements/         ← Layer 2: non-peptide nutritional adjuncts (CoQ10, NAD, …)
    β”œβ”€β”€ Compounds/           ← Layer 2: non-peptide therapeutics (MK-677, orforglipron, SERMs, …)
    β”œβ”€β”€ Behaviors/           ← Layer 2: lifestyle interventions (sauna, cold, Zone 2, HIIT, fasting, sleep, sunlight)
    β”œβ”€β”€ Stacks/              ← Layer 3: applied protocols (peptides + supplements + behaviors)
    β”œβ”€β”€ Methods/             ← cross-cutting reference methods (animalβ†’human dose conversion, …)
    └── Bases/
        β”œβ”€β”€ Articles.base
        β”œβ”€β”€ Peptides.base
        └── Papers.base

Channel output stays where it already lives: 01 - Distribution/Content/ (Layer 3). The knowledge base feeds it; they do not merge.

5. Note Schemas (frontmatter)

Aligned to Data Model & Conventions golden rules: frontmatter is the data, filename = ID, dates are ISO, money is numeric, cross-links are wikilinks, tag drives the Base.

Article β€” #kb-article

type: kb-article
id: ART-bpc-157-complete-guide-2026
title: "BPC-157 Complete Guide 2026: Benefits & Dosing"
slug: bpc-157-complete-guide-2026
source_url: https://indexalabs.com/blog/bpc-157-complete-guide-2026-benefits-dosing
published: 2026-03-10
cohort: seo-2026            # or product-2025
pillar: education           # education | product | social-proof | promo | research
primary_peptide: "[[BPC-157]]"
peptides: ["[[BPC-157]]"]
mechanisms: ["[[Angiogenesis]]", "[[Cytoprotection]]"]
papers: ["[[PMID-12345678]]"]
products: [bpc-157]         # storefront slugs (Supabase), NOT vault notes β€” plain strings, never wikilinks
target_keyword: "bpc-157 dosing"
scrape_status: mirrored     # pending | mirrored | extracted | verified
content_hash: ""            # sha256 of source body, for re-scrape diffing
tags: [kb-article]

Peptide β€” #kb-peptide (the hub note)

type: kb-peptide
id: BPC-157
title: "BPC-157"
aliases: ["Body Protection Compound-157", "PL 14736"]
category: tissue-repair     # matches site category taxonomy
sequence: "GEPPPGKPADDAGLV"
molar_mass: 1419.5
half_life: "~4 hours (subcutaneous, est.)"
mechanisms: ["[[Angiogenesis]]", "[[Cytoprotection]]"]
stacks_with: ["[[TB-500]]"]
papers: ["[[PMID-12345678]]"]
product: bpc-157            # storefront slug (Supabase), NOT a vault note β€” plain string
forms: []                   # extra product slugs that are delivery-forms of this compound
is_blend: false             # true for blends; then list components as wikilinks
components: []              # for blends only: ["[[BPC-157]]", "[[TB-500]]"]
evidence_tier: B            # S | A | B | C | D (Research-pillar tier list)
research_use_only: true
tags: [kb-peptide]

Mechanism β€” #kb-mechanism

type: kb-mechanism
id: Angiogenesis
title: "Angiogenesis"
domain: vascular            # vascular | metabolic | neuro | immune | hormonal | dermal
peptides: ["[[BPC-157]]", "[[GHK-Cu]]"]
papers: ["[[PMID-12345678]]"]
tags: [kb-mechanism]

Paper β€” #kb-paper (evidence node)

type: kb-paper
id: PMID-12345678
title: "Pentadecapeptide BPC 157 and tendon healing"
authors: "Sikiric P, et al."
year: 2014
journal: "J Orthop Res"
doi: "10.1002/jor.xxxxx"
pmid: 12345678
url: https://pubmed.ncbi.nlm.nih.gov/12345678/
model: animal              # in-vitro | animal | human-rct | human-obs | review
peptides: ["[[BPC-157]]"]
mechanisms: ["[[Angiogenesis]]"]
evidence_weight: moderate  # strong | moderate | weak | anecdotal
tags: [kb-paper]

Concept β€” #kb-concept

type: kb-concept
id: Reconstitution
title: "Reconstitution"
related: ["[[Bacteriostatic Water]]", "[[Storage & Handling]]"]
tags: [kb-concept]

Enrichment fields (added over time, never overwriting source): atomic notes carry free-text body sections for ## Anecdotes, ## Physician protocols, ## Practical notes, each line provenance-tagged (see Β§9).

6. Naming & ID Conventions

Extends the existing ID table in Data Model & Conventions:

EntityPatternExample
KB ArticleART-<slug> (filename = article title)ART-bpc-157-complete-guide-2026
Peptide<COMPOUND> (filename = common name)BPC-157
Mechanism<Name>Angiogenesis
PaperPMID-<id> or DOI-<id>PMID-12345678
Concept<Name>Reconstitution

Rules: filename matches the human-readable title so [[BPC-157]] resolves naturally; id in frontmatter is the stable machine key for Supabase. Slugs mirror the live URL exactly so we can round-trip to the site.

7. Interlinking & SEO Policy

The link graph is the SEO asset. Rules:

  1. Every article links to its primary peptide and to every mechanism/paper it cites. No orphan articles.
  2. Hub-and-spoke: Peptide notes are hubs. Articles, papers, and mechanisms spoke into them. A peptide hub should be reachable in ≀2 hops from any related note.
  3. Reciprocal links: if an article links a peptide, the peptide’s peptides/related list need not list every article (Bases derive that), but mechanism↔peptide and peptide↔paper links are maintained on both ends.
  4. Cluster topology mirrors site categories (GLP-1/Metabolic, Growth Hormone, Tissue Repair, Nootropics, etc.) so internal-link clusters map to the site’s category silos β€” this is what compounds ranking.
  5. Comparison articles (β€œX vs Y”, β€œbest peptides for…”) must link both/all peptides and the relevant mechanism β€” these are the highest-value link hubs.
  6. Anchor text = descriptive, never β€œclick here”. Wikilink display text should read as the target keyword where natural.
  7. No dead links. A [[wikilink]] to a not-yet-created atomic note is allowed (it marks a stub to fill) but stubs are tracked and burned down before ingestion.

8. Scraping Methodology & Policy

Source of record: the live site sitemap (indexalabs.com/sitemap.xml) β€” already confirmed to list 82 /blog/ URLs plus resources and product pages. The sitemap is the canonical work list, not the rendered /blog index.

The 82 split into two cohorts (drives cohort field and review priority):

  • product-2025 (~40 posts, lastmod 2025-01) β€” one-per-product research guides aligned to SKUs.
  • seo-2026 (~42 posts, lastmod 2026-03) β€” comparison, β€œbest-of”, dosing, and delivery-method articles built for search.

Method (per article):

  1. Pull the URL list from the sitemap; write a manifest (Bases/Articles.base seeded as scrape_status: pending).
  2. Fetch each article. The site is partly client-rendered, so the pipeline must render JS (browser-based fetch) rather than raw HTML when the body is empty.
  3. Convert HTML body β†’ clean Markdown (headings, lists, tables preserved; nav/footer/boilerplate stripped).
  4. Write the Article note (Layer 1) with full frontmatter, source_url, and a content_hash of the source body.
  5. Set scrape_status: mirrored.

Extraction (Layer 1 β†’ Layer 2), per article:

  1. Identify the peptide(s), mechanism(s), and any cited papers. Create/append the atomic notes (don’t duplicate β€” link if they exist).
  2. Replace inline factual claims in the article with links to the atomic notes that own them.
  3. Set scrape_status: extracted, then verified after the Β§9 review.

Re-scrape policy: on a scheduled cadence, re-fetch and compare content_hash. If changed, flag the article for re-mirror + re-extract so the base never drifts from the live site.

Politeness & integrity: scrape our own site only, throttle requests, preserve published dates, never invent content that isn’t on the page. Mirrors are faithful; interpretation happens only in Layer 2 with citations.

9. Provenance, Enrichment & Citation Policy

The base’s value is trust, so every claim is traceable.

  • Three provenance classes, tagged on each enrichment line/section:
    • [indexa] β€” already published by us (from a mirrored article).
    • [literature] β€” backed by a [[PMID-…]] paper note (Research pillar).
    • [anecdote] / [protocol] β€” community report or physician protocol; explicitly marked as such, never presented as established fact.
  • No unsourced claims enter Layer 2. If it can’t be tagged to one of the above, it doesn’t get written.
  • Evidence tiers (evidence_tier Sβ†’D on peptides, evidence_weight on papers) power the β€œMost Researched Peptides Tier List” Research-pillar asset and let channels signal confidence honestly.
  • Stack backing (mandatory): every synergy, dosing, timing, intake, and cycling claim in a kb-stack must trace to a [[paper]] (literature) or a named applied-protocol source ([[transcript]], e.g. a physician). The stack’s backed_by: frontmatter lists them, and the ## Evidence backing section maps claims β†’ sources. A stack with unbacked claims stays draft, never verified. Dosing/timing figures are recorded as cited research / applied-protocol information under research-use framing β€” never as personalised medical advice.
  • Compliance guardrail (non-negotiable): every peptide/article note carries research_use_only: true, and no note makes medical-treatment or dosing-for-humans claims. Framing is research/educational throughout β€” consistent with the site’s own β€œresearch and educational purposes only” stance. This guardrail propagates to Layer 3.

10. Supabase Ingestion Model

Obsidian is the source of truth (write side); Supabase is the read model the site/app queries. The forward pipeline already exists β€” it does not need to be built.

What’s already live (project gqkhhfuafflcfkjczsxg):

  • The ingest-content Edge Function (/functions/v1/ingest-content, auth via scoped x-ingest-key) maps note frontmatter β†’ the live storefront tables: guides (+ guide_sections, guide_products) and products (+ research/variants/prices/images/tags). Upserts by slug; everything lands published = false (draft) until reviewed in Studio.
  • 86 published guides plus guide_sections are the canonical blog content. This KB’s Layer-1 Articles folder was backfilled one-time from those rows (Supabase β†’ vault) so the vault becomes the forward author-source. From here, edits flow vault β†’ ingest-content β†’ Supabase draft β†’ publish.

The gap β€” the atomic graph has no DB home yet. The live schema stores articles (guides) and products, but not the Layer-2 atomic notes (peptides, mechanisms, papers, concepts). Decision needed (see Β§14): either (a) add kb_peptides / kb_mechanisms / kb_papers / kb_concepts + join tables and extend ingest-content to populate them, or (b) keep the atomic graph vault-side only and surface it to customers via RAG/embeddings rather than relational tables.

Direction stays one-way: vault β†’ Supabase (via ingest-content). Supabase rows are never hand-authored as the source; the vault is. The one exception was this initial backfill, now complete.

11. Knowledge β†’ Content Manipulation Layer (Layer 3)

This is the bridge to TikTok/Threads, governed by Content β€” Playbook. The knowledge base supplies what’s true; the playbook supplies how each channel speaks.

  • One-way borrow. A content post references the atomic notes it draws from (source_notes: ["[[BPC-157]]"]) so we can trace any claim back to the base. Posts never edit the base.
  • Channel tone is applied on extraction, not stored in the base. The peptide note is neutral and evidence-led; the TikTok script is punchy; the Threads alias post is community-voiced. Same fact, three packagings.
  • Threads alias rule (from memory, preserved): Threads posts run under disposable alias brands, never as INDEXA β€” but the facts still come from the INDEXA base. The base is brand-neutral truth; the alias is just a delivery vehicle.
  • Research pillar pulls directly from kb-paper notes and the tier list β€” this is the trust engine for the knowledge-minded Threads audience.
  • Guardrail inheritance: research_use_only and the no-medical-claims rule carry into every derived asset automatically.

12. Governance β€” Source of Truth Rules

  1. Single source of truth: Layer 1 + Layer 2 in Obsidian. The live site and Supabase are projections; social posts are derivatives.
  2. One thing, one note (inherited golden rule). Facts are written once in the atomic note that owns them.
  3. Frontmatter is the data; the body is context + provenance-tagged enrichment.
  4. No fact without provenance (Β§9). No medical/treatment claims, ever.
  5. Mirrors stay faithful; interpretation stays cited. Editorialising happens only in Layer 2, against papers.
  6. Changes flow downhill: vault β†’ Supabase β†’ site/app; vault β†’ playbook β†’ channels. Never upward.
  7. Stubs are debt: unfilled [[links]] are tracked and burned down before a cluster is marked verified.

13. Build Roadmap

PhaseDeliverableStatus
0. StrategyThis document, approvedβœ… done
1. ScaffoldFolder tree + 5 note templates + 3 Bases (manifest seeded from sitemap, 86 rows pending)βœ… done
2. MirrorAll 86 articles backfilled from Supabase guides/guide_sections as Layer-1 notes (mirrored, 455 sections)βœ… done
3. ExtractPeptide/Mechanism/Paper/Concept atomic notes created; articles re-linked (extracted). Seed papers from references_raw.β¬… next
4. VerifyProvenance + compliance pass; stubs burned down (verified)
5. Ingest (forward)ingest-content Edge Function already live (vault β†’ Supabase drafts). Extend to atomic graph only if Β§14.6 picks relational tables.partial β€” forward pipeline exists
6. OperateRe-sync cadence + content-borrow workflow into Layer 3

14. Open Decisions β€” need your call before Phase 1

  1. Atomic depth for the first pass. Mirror all 82 first (Phase 2 complete) then extract, or do mirror+extract article-by-article? (Recommend: mirror all 82 first β€” faster to a complete Layer 1, cleaner dedupe in extraction.)
  2. Papers scope. Do articles already cite specific PubMed IDs we can harvest, or do we build kb-paper notes from scratch during extraction? (Changes Phase 3 effort significantly.)
  3. Supabase + GitHub access. Share the Supabase project (schema/keys) and the vault’s git repo if you want CI-driven ingestion β€” otherwise I’ll spec a manual export script.
  4. Bases vs. backlinks for reverse edges. Confirm we rely on Obsidian Bases to derive β€œarticles mentioning peptide X” rather than maintaining reverse lists by hand (recommended β€” less drift).
  5. Resource pages. Include all six site resources (dosage, reconstitution, calculator, storage, payment, dictionary) as kb-resource, or only the science-bearing ones (exclude payment)? (Recommend: exclude payment-guide β€” not knowledge.)
  6. Atomic-graph DB home. The live schema has guides + products but no home for peptide/mechanism/paper/concept notes. Either (a) add kb_* tables + join tables and extend ingest-content, or (b) keep the atomic graph vault-side and serve it to customers via RAG/embeddings. (Recommend: start with (b) β€” the graph is for learning/retrieval, not storefront rows; relational tables can come later if a feature needs them.)

Phases 0–2 are complete: strategy approved, vault scaffolded, all 86 articles mirrored from Supabase. Next is Phase 3 β€” extracting the atomic peptide/mechanism/paper graph and re-linking the articles into it.

Graph View