Back office ops · Production

Engineering the Memory Layer for an AI Agent to Navigate Large-Scale Event Data on ApertureDB

The problem

Hundreds of conference talks spanning multiple years had no good way to search without clicking through many pages; traditional keyword search failed for complex semantic queries requiring metadata filtering and entity relationships.

First attempt

Traditional keyword search and flat table structures could not handle multi-part queries combining semantic search, metadata filtering, and entity joins without complex application-level orchestration, and speaker data stored as comma-separated strings made targeted queries unreliable.

Workflow diagram · grounded in source
1
Raw conference data intake
trigger
“The source data consists of talk submissions from MLOps World and GenAI World conferences (2022-24). After deduplication and filtering for talks with available YouTube recordings, the working dataset has 280 unique talks from 263 speaker…”
2
YouTube metadata enrichment
integration
“We can enrich the talk information further via Apify on YouTube with additional metadata: view counts, publish dates, and timestamped transcript segments that would later enable precise video linking from search results.”
3
Entity graph ingestion
integration
“338 Person entities and 373 TalkHasSpeaker connections (some speakers presented multiple talks across conferences). ApertureDB's if_not_found clause on both entities and connections ensures complete idempotency”
4
Transcript chunking with timestamps
ai_action
“the preserved timestamps (start_sec, end_sec) enable a powerful UX capability: search results can deep-link directly to the relevant moment in the source video, transforming semantic search from "this talk mentions X" to "this exact 30-s…”
5
Vector ingestion into ApertureDB
integration
“Processing 278 talks yielded 16,887 transcript chunks. ApertureDB's ParallelLoader completed ingestion in approximately 4 minutes at 71.5 items/second, including both embedding storage and connection creation in atomic transactions.”
6
AI agent natural language query
ai_action
“Our tool-equipped AI Query Agent is able to navigate rich interconnected data stored in ApertureDB, an AI-native multimodal vector-graph database, to answer natural language queries effectively and accurately”
Reported outcome

A tool-equipped AI Query Agent was built that navigates rich interconnected data in ApertureDB to answer natural language queries effectively and accurately, backed by 280 talks, 338 person entities, and 16,887 transcript chunk embeddings with video deep-link capability.

Reported metrics
Unique conference talks in dataset280
Unique speakers263
Transcript chunk embeddings16,887
Ingestion speed71.5 items/second
Show all 7 reported metrics
unique conference talks in dataset280
unique speakers263
transcript chunk embeddings16,887
ingestion speed71.5 items/second
ingestion durationapproximately 4 minutes
Person entities created338
TalkHasSpeaker connections373
Reported stack
ApertureDBEmbeddingGemmaGoogle ColabApifyLangGraphYouTube
Source
https://mlops.community/blog/engineering-the-memory-layer-for-an-ai-agent-to-navigate-large-scale-event-data
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

A tool-equipped AI Query Agent was built that navigates rich interconnected data in ApertureDB to answer natural language queries effectively and accurately, backed by 280 talks, 338 person entities, and 16,887 transc…

What tools did this team use?

ApertureDB, EmbeddingGemma, Google Colab, Apify, LangGraph, YouTube.

What results were reported?

Unique conference talks in dataset: 280; Unique speakers: 263; Transcript chunk embeddings: 16,887; Ingestion speed: 71.5 items/second (source-reported, not independently verified).

What failed first in this deployment?

Traditional keyword search and flat table structures could not handle multi-part queries combining semantic search, metadata filtering, and entity joins without complex application-level orchestration, and speaker dat…

How is this back office ops AI workflow structured?

Raw conference data intake → YouTube metadata enrichment → Entity graph ingestion → Transcript chunking with timestamps → Vector ingestion into ApertureDB → AI agent natural language query.

Related back office ops cases