How it works
The pipeline behind the app — collect, enrich, store, ask — and what the research underneath it adds.
- Collect
Data ingestion & the Instagram source
Authorised Graph API access or ethical collection tooling, handling authentication, rate limiting and PII anonymisation. Raw output: posts, reels, stories and comments with timestamps, likes, saves, shares, hashtags, follower ties and author information.
In this build: This build ships an anonymised demonstration corpus of 28 posts and 20 follower ties so the pipeline is reproducible without platform credentials; your own account can be linked on My corpus.
- Enrich
AI-augmented metadata extraction
Text pre-processing and anonymisation, followed by named entity and concept extraction, topic modelling, sentiment and emotion scoring, and network graph mapping for influencers and community clusters.
In this build: Implemented as deterministic in-browser surrogates for spaCy, BERTopic and RoBERTa, plus label-propagation community detection over the follower graph.
- Store
The generative knowledge core
A vector index holding semantic embeddings of enriched chunks, a graph store holding relational metadata, and a retrieval engine combining semantic search with graph traversal.
In this build: A 96-dimension hashed embedding index with cosine retrieval, boosted by author influence and hashtag overlap.
- Ask
Answers and knowledge consumption
A retrieval-augmented language model synthesises answers for decision makers, and the results are consumed through the overview, corpus inspector, network analysis and query interface.
In this build: Synthesis runs server-side against a hosted language model, returning citations to post identifiers that open in the evidence panel.
Knowledge facets captured per record
| Who | Author identity, category, follower ties and community membership |
| What | Caption text, entities, modelled topic, hashtags and media type |
| When | Post timestamp, collection window and sentiment trend position |
| Where | Location tag and the spatial spread of the conversation |
| How | Engagement mechanics: likes, comments, saves, shares, amplification |
| Why | Motive inferred from topic, emotion and comment discourse |
From tacit to explicit knowledge
Posting behaviour — composing a caption, selecting hashtags, tagging a location, responding in comment threads — externalises judgement that would otherwise stay tacit. The enrichment layer codifies that behaviour as structured metadata, the vector and graph stores combine it across accounts, and the generative layer internalises it again as a cited brief a decision maker can act on.
Limitation acknowledged in the study: metadata can only capture the explicit and implicit strata. Fully tacit expertise resists codification, so generated output is presented as evidence-linked synthesis rather than automated decision making.