AI Platform · Retrieval Fabric
An answer is only as good as the passage behind it
Hybrid retrieval over five indexes, reranked by a cross-encoder, then grounded claim by claim. Anything that cannot be traced to a passage is deleted before the answer leaves the agent.
Indexed chunks
10.8M
Mean recall@8
92%
measured weekly on labelled queries
Grounded claim rate
100%
ungrounded sentences are removed
Retrieval p95
510ms
rewrite → rerank → ground
Try the retrieval path
A real query walks the five stages with its latency budget
- 1Query rewrite120ms
Resolves 'it', 'that stock' and vernacular tickers against the member's last 5 turns.
- 2Hybrid retrieve210ms
BM25 + vector, reciprocal-rank fused, k=48 before rerank.
- 3Rerank180ms
Cross-encoder keeps 8 passages; anything under 0.42 is dropped, not padded.
- 4Ground & cite90ms
Every claim maps to a passage id or the sentence is deleted before generation ends.
- 5Compliance pass260ms
Guard reads the drafted answer, not the prompt, and can veto.
Indexes
Chunking is chosen per corpus, not globally
| Index | Documents | Chunks | Chunking strategy | Recall@8 | Freshness | Dims |
|---|---|---|---|---|---|---|
| Exchange filings & annual reports | 1,84,000 | 64,20,000 | Section-aware, 900 tokens, 12% overlap | 93% | 8 min | 3072 |
| Broker & sell-side research | 41,200 | 11,80,000 | Table-preserving, 1200 tokens | 89% | 1 hour | 3072 |
| SEBI circulars & regulation | 9,700 | 2,12,000 | Clause-level, no overlap | 96% | 24 hours | 3072 |
| StockGro product & help content | 3,100 | 41,000 | Heading split, 600 tokens | 94% | 5 min | 1536 |
| Member conversation memory | 8,12,000 | 29,40,000 | Turn-pair with intent labels | 87% | live | 1536 |
Recall by index
Regulation is the most exact corpus, memory the least
What we refuse to retrieve
Corpus hygiene is a compliance control
- Telegram and WhatsApp tip channels — the source of most manipulation narratives.
- Unverified social claims about unlisted or SME scrips.
- Any document without a resolvable publisher and date.
- Member conversation memory when the purpose token does not name that member.