Overview

Every item is meant to be findable by people, place, event, and theme, across English and Persian. Machine learning accelerates labeling, but sensitive merges and conflict resolution stay with reviewers so the catalog does not silently absorb mistakes. Items from link intake and similar sources do not enter the public export until a curator marks them verified.

Ingestion is context-aware: each link or post is normalized, then matched to domain-specific prompt packs and to retrieval hits from the published catalog before vision tagging runs. That keeps labeling consistent with source type (Telegram channel, memorial roster, X post, open-web image) and with what the archive already documents.

%%{init: {"flowchart": {"padding": 20, "nodeSpacing": 50, "rankSpacing": 58}}}%%
flowchart TB
  subgraph sources [Sources]
    TG["Grassroots<br/>channels"]
    MR["Memorial<br/>rosters"]
    SOC["Hashtag<br/>harvest"]
    LI["Link<br/>intake"]
  end
  subgraph corePipe [Core pipeline]
    IN["Ingest"]
    ST["Catalog"]
    RAG["RAG<br/>index"]
    ROUTE["Prompt<br/>routing"]
    TAG["Vision<br/>tagger"]
    FACE["Face<br/>match"]
  end
  subgraph qa [Quality assurance]
    CUR["Curator<br/>review"]
    DUP["Dedup<br/>review"]
    NAMES["Name<br/>review"]
    EDIT["Overrides"]
  end
  subgraph out [Publication]
    PKG["Export"]
    WEBP["Static<br/>site"]
    SRCH["Search<br/>index"]
    GRAPH["Graph<br/>beta"]
  end
  TG --> IN
  MR --> IN
  SOC --> IN
  LI --> IN
  IN --> ST
  ST --> RAG
  ST --> ROUTE
  RAG --> ROUTE
  ROUTE --> TAG
  TAG --> FACE
  FACE --> CUR
  CUR --> DUP
  CUR --> NAMES
  DUP --> EDIT
  NAMES --> EDIT
  EDIT --> PKG --> WEBP
  PKG --> SRCH
  PKG --> GRAPH

Sources and connectors

%%{init: {"flowchart": {"padding": 20, "nodeSpacing": 50, "rankSpacing": 58}}}%%
flowchart TB
  subgraph tel [Telegram]
    CH["Public<br/>channels"]
    CAP["Captions"]
    BOT["Discovery<br/>intake"]
  end
  subgraph mem [Memorial feeds]
    RO["Roster data"]
    PT["Portraits"]
  end
  subgraph cross [Social and web]
    HASH["Hashtag<br/>search"]
    XPM["Posts<br/>from X"]
    WEB["Web pages"]
  end
  CH --> CAP
  RO --> PT
  BOT --> urlDec["URL<br/>normalize"]
  urlDec --> linkQ["Link<br/>inbox"]
  CAP --> catalog["Catalog"]
  PT --> catalog
  HASH --> XPM
  HASH --> WEB
  XPM --> ctx["Cross-channel<br/>context"]
  WEB --> ctx
  ctx --> catalog
  linkQ --> catalog

Storage

AI enrichment

Google Gemini powers vision-first passes over photos and sampled video frames: full descriptions in English and Persian, semantic tags, people and places, and event hints, including detection of جاویدنام (Javidnam) memorial content where appropriate. Separate text calls condense many captions into structured memorial profiles (age, death context, hometown, and narrative bio) so the جاویدنامان index can show more than a single photo. Evidence from the social and web harvest layer is folded into that same enrichment pass as supporting context.

Context routing works like selecting specialized editorial skills for each item. After a URL is decoded (Telegram, X, or direct image), the pipeline records connector, channel, and caption on the catalog row. A router then layers the right instruction packs: a shared bilingual tagging base, plus add-ons for link-inbox Telegram or X, memorial rosters, era-specific Twitter rules, or named channels with their own naming and event conventions. The result is one composed prompt per item, not a single generic template.

In parallel, a RAG-style index built from published export text is queried on ingest. Top matching chunks (existing items, memorial rows, related captions) support duplicate checks, link-inbox verification, and reviewer context. Retrieval and prompt packs address the same goal, grounding new material in prior catalog evidence, through vector search and through explicit source-specific rules.

The vision tagger receives that composed prompt, the media (photo or sampled video frames), and optional curator hints. It returns structured English and Persian descriptions, tags, people, places, and event fields that are merged with caption-derived facts and post-processing rules before anything is stored.

%%{init: {"flowchart": {"padding": 20, "nodeSpacing": 50, "rankSpacing": 58}}}%%
flowchart TB
  subgraph intake [Intake]
    URL["URL or<br/>caption"]
    DEC["Normalize<br/>and fetch"]
    META["Item<br/>metadata"]
  end
  subgraph context [Context assembly]
    RAG["RAG top-k<br/>chunks"]
    PACK["Prompts<br/>Skills"]
    MERGE["Composed<br/>prompt"]
  end
  subgraph model [Vision tagging]
    GEM["Gemini<br/>vision pass"]
    OUT["Labels<br/>and text"]
  end
  URL --> DEC --> META
  META --> PACK
  META --> RAG
  PACK --> MERGE
  RAG --> MERGE
  MERGE --> GEM --> OUT

AI output is treated as assistive: it speeds curation but does not bypass editorial judgment on politically or emotionally charged material, and it does not replace curator sign-off on sensitive intake.

Quality, deduplication, and naming

%%{init: {"flowchart": {"padding": 20, "nodeSpacing": 50, "rankSpacing": 58}}}%%
flowchart TB
  subgraph inboxTrack [Link inbox]
    VER["Auto<br/>checks"]
    FD["Face<br/>detect"]
    FM["Gallery<br/>match"]
    RV0["Human<br/>review"]
    VER --> RV0
    FD --> FM --> RV0
  end
  subgraph mediaTrack [Near duplicates]
    PH["pHash"]
    EM["CLIP<br/>embed"]
    VX["FAISS<br/>search"]
    RV1["Human<br/>review"]
    PH --> cand["Candidates"]
    EM --> cand
    VX --> cand
    cand --> RV1
  end
  subgraph nameTrack [Names]
    FZ["Fuzzy<br/>match"]
    RV2["Canonical<br/>pick"]
    FZ --> RV2
  end
  subgraph editTrack [Editorial]
    MAN["Hand<br/>edits"]
  end
  RV0 --> canonCat["Catalog"]
  RV1 --> canonCat
  RV2 --> canonCat
  MAN --> canonCat
  canonCat --> pubSite["Public site"]

Publication

The public site is a static build: pre-rendered HTML, no server-side database in front of visitors. Browsing runs through the archive grid, tag index, event index, item detail pages, and the جاویدنامان memorial roll with profile pages. Bilingual fields in the export drive English and Persian labels where both exist.

Full-text search uses an index generated from exported content and hosted on CDN alongside media, separate from the HTML deploy so search can be refreshed without rebuilding every page.

A graph network (beta) is exported from co-occurrence and shared-item links between people, places, and archived media. Explore views on جاویدنامان profiles show related names and connections; the feature is experimental and may change as the catalog grows.

Deployments flow from the project repository to edge-hosted static pages, with media served from object storage. Curated JSON and HTML are committed after human review passes complete, keeping the live site aligned with verified catalog state.

Glossary

Javidnam / جاویدنام

Eternal-name memorial framing for people killed in connection with the protest movement; items and profiles can carry this designation.

Link inbox

Intake queue for new URLs and media awaiting curator review before they are included in the public export.

Curator verified

Human approval flag; required for link-inbox-class and similar intake sources in the published build.

Domain prompt pack

Source-specific instruction text (channel, connector, era) composed with the shared tagging base before a vision model run.

Context routing

Rule-based selection of prompt packs from catalog metadata after URL normalize and fetch, analogous to choosing an editorial skill for the item.

RAG index (ingestion)

Embedded text chunks from the published catalog, retrieved during ingest to supply neighborhood context for verification and review.

Perceptual hash (pHash)

A fingerprint resilient to minor editing; useful to cluster visually similar images without byte-identical files.

Embedding (CLIP-class)

A numeric vector summarizing image content so “same scene” pairs can be found even when pixels differ strongly.

Vector index (FAISS-style)

Approximate nearest-neighbor search over millions of vectors to retrieve similar images quickly.

Face gallery match

Detected faces compared to embedded portraits in the catalog for verification hints and memorial linkage.

Graph network (beta)

Exported links between people and shared archive items, shown on Explore memorial views; still evolving.