- 1Tech Note
The questions were deciding the scope of the answers (Part 1)
With only shipping and returns questions, the exception approval process in the source never entered the graph. Moving questions to evaluation.
- 2Tech Note
What we truncated to save context was the evidence (Part 2)
We added multi-turn search to follow relations. Context grew, so we cut tool results to 500 characters, and long sources never reached synthesis.
- 3Tech Note
We had put a multi-hour build in a cache (Part 3)
As builds stretched into hours we replicated job state into a shared cache. When it expired we could not recover cancellation and resume state.
- 4Tech Note
Triple counts told us nothing about graph quality (Part 4)
The same pipeline produced 250,000 triples on one input and 1,000 on another. Both had problems, running in opposite directions.
- 5Tech Note
Cleaning up the graph deleted 1,500 healthy classes (Part 5)
Auto-deleting classes short on instances and relations removed about 1,500 healthy ones. Post-processing judged what extraction had not yet built.
- 6Tech Note
The A/B calling it slow and wrong was measuring an empty graph (Part 6)
An A/B said multi-turn search was slower and less accurate. The evaluation was querying an empty graph and answering from vector search alone.
- 7Tech Note
Most of the nodes we painted red were not evidence (Part 7)
With retrieval quality raised, verifying on screen showed far more nodes highlighted than there was evidence for. What the highlight really proves.
- 8Tech Note
Document dedup was merging structured tables (Part 8)
We built ten CSV tables and the post-processed graph held four classes. A rule for tidying document concepts had merged schema identifiers.
- 9Tech Note
The job reported success and the graph was empty (Part 9)
Switching to a small local model left batches unfinished, and jobs recorded as successful left an empty graph. Truncation arrived as empty results.
- 10Tech Note
URIs built from names merged two people into one (Part 10)
We used the database's types and keys directly. Identifiers made from display names merge namesakes and split renamed rows.
Try XGEN free for 15 days — nothing to install, runs in your browserStart free trial
