Services
Retrieval-augmented generation built on measurement: hybrid BM25 plus dense retrieval, reranking that actually moves recall, eval sets graded by your subject-matter experts, and permission-aware indexes that respect tenant and ACL boundaries. We size vector stores for 18 months of corpus growth, pick embeddings on data not vendor decks, and refuse to ship a RAG that has not survived a regression eval. Audits from 8,500 EUR, working pilots up to 10K documents from 35,000 EUR, production retainers from 14,000 EUR per month.
Most enterprise RAG systems fail at the same chokepoint: retrieval. The LLM is fine. The prompt is fine. But recall@5 sits at 40 percent, the user never sees the right chunk, and the answer is plausibly wrong. We start with corpus profiling and a 200 to 500-question eval set graded by your experts — not vibes. We benchmark embeddings, chunk sizes, and retrieval strategies as parameters, not opinions. Hybrid retrieval with reranking is the default because dense alone misses SKU and clause-number queries. Permission-aware filtering is enforced at query time, because misconfigured RAG is the most common cause of accidental data exposure. By week 8, you have a RAG you can defend to security review.
Connectors for SharePoint, Confluence, Google Drive, S3, Slack, Notion, and database extracts. Recursive splitters tuned to your document distribution — legal prose, technical docs with code, transcripts — with overlap calibrated on your eval set.
Benchmark of OpenAI, Cohere multilingual, and BAAI bge on your eval set. We pick on measured recall@k and cost per million tokens at your corpus size, not on the model leaderboard from last quarter.
pgvector for under a few million vectors when ops simplicity wins, Qdrant or Weaviate for self-hosted at scale, Pinecone for managed, OpenSearch when hybrid is already your search backbone. Sized for 18 months of growth.
Reciprocal rank fusion of BM25 and dense retrieval so exact-match queries (SKUs, clause numbers, ticket IDs) and paraphrase queries both work. Tuned on your eval set, not on a default mix.
Cross-encoder reranking with Cohere Rerank 3 or bge-reranker-v2-m3, lifting recall@5 by 15 to 30 percent over dense-only. Faithfulness and answer-relevance rubrics running in CI on every prompt or index change.
ACL metadata attached at ingest, enforced at query time. Per-tenant index isolation for high-sensitivity corpora. Audited explicitly because misconfigured RAG is the single most common source of accidental data exposure.
Weeks 1–2: corpus profile, current-state assessment, 200 to 500-question eval set built with your subject-matter experts, written architecture recommendation with ADRs.
Weeks 3–4: connectors, chunking strategy, embedding benchmark, vector store stood up, ACL metadata mapped, initial index built and backfilled.
Weeks 5–7: hybrid retrieval tuned, reranking integrated, generation prompt versioned in git, eval harness running in CI, customer-zero deployment behind a flag.
Week 8+: observability live (Phoenix, TruLens, LangSmith), permission audit signed off, runbook written, your team trained on adding corpora and expanding the eval set.
Two weeks. Corpus profile, eval-set design, current-state assessment if you have an existing RAG, architecture recommendation with ADRs. Best when you do not yet know whether your RAG is broken or the prompt is. 8,500 EUR fixed.
6 to 8 weeks for up to 10,000 documents. Ingestion pipeline, vector store, hybrid retrieval, reranking, eval harness in CI, ACL enforcement, customer-zero deployment. 35,000 EUR fixed.
Monthly. Corpus expansion, embedding upgrades, eval growth, retrieval tuning, on-call for RAG-specific incidents. Best after pilot ships and corpus is still growing. From 14,000 EUR/month.
All engagements start with a mutual NDA, IP assignment, and a DPA. Three-month minimum on the production retainer, month-to-month thereafter with 30 days notice.
Native iOS and Android e-signature clients with a Symfony + React CRM for a cross-border law firm — KYC onboarding and a defensible evidence trail for US & EU matters.
B2B e-commerce and product configurator for a global polymer manufacturer with multi-region pricing, stock and dealer workflows.
Production social platform — App Store + Google Play, live across the US and EU — with geo Radar, encrypted messaging and a virtual economy.
GDPR-aligned · ISO 27001 ready · SOC 2 Type II in progress · HIPAA-capable · CCPA-acknowledged
Every chunk size, every embedding, every retrieval mix is picked on recall@k on your eval set. No best practices imported from a blog post. We can show you the numbers behind every choice.
ACL enforcement at query time is built in from week one, not bolted on at security review. We audit it explicitly because misconfigured RAG is the single most common GenAI security incident.
Our RAG leads have run search relevance and embedding pipelines before LLMs were the answer. They argue about recall@5 and reranker latency, not about which paper dropped last week.
We treat RAG as a search system with a generation step on top — not the other way around. Retrieval is where the value lives and where the bugs hide.
Chunking is empirical, not theoretical. We profile your corpus first: median document length, paragraph distribution, section structure, table density, code-block presence. For dense prose (legal, policies) we typically start at 400 to 600 tokens with 15 percent overlap; for technical docs with code we move to recursive splitters that respect code-block and heading boundaries; for transcripts we chunk on speaker turns plus token cap. Then we run a retrieval eval across three chunk sizes and pick the one that maximises recall@10 on the actual eval set. No rule of thumb beats measurement on your data.
Embeddings: we benchmark OpenAI text-embedding-3-large, Cohere embed-multilingual-v3, and BAAI bge-large on your eval set; the winner depends on language mix and domain. Vector store: pgvector if you already have Postgres and corpus is under a few million vectors (operational simplicity wins); Qdrant or Weaviate for self-hosted at scale with metadata filtering; Pinecone when ops budget is tight and you want managed; OpenSearch when hybrid BM25 plus dense is already your search backbone. We size the index for 18 months of corpus growth, not today's snapshot.
Almost always. Pure dense retrieval misses exact-match queries (product SKUs, contract clause numbers, ticket IDs); pure BM25 misses paraphrase. We combine BM25 and dense with reciprocal rank fusion, then rerank the top 50 with a cross-encoder (Cohere Rerank 3 or bge-reranker-v2-m3) down to the top 5 to 10 that go into the prompt. Reranking typically lifts recall@5 by 15 to 30 percent over dense-only on enterprise corpora — we measure it on your eval set before recommending it for production.
Three layers of eval. Retrieval: recall@k, MRR, and a relevance-graded judge on a 200 to 500-question eval set built with your subject-matter experts. Generation: faithfulness (Ragas) plus an answer-relevance rubric so we catch hallucinations and topic drift. End-to-end: a human-graded sample of 50 to 100 production-like queries each week during pilot. All three run in CI on every prompt or index change. Any tier-1 regression blocks merge. Production observability through Phoenix, TruLens, or LangSmith logs every retrieval and generation for offline analysis.
Yes, and it is the single most common requirement we see. Permission-aware retrieval enforces ACLs at query time, not at index time. We attach access metadata (user IDs, group IDs, tenant IDs, sensitivity labels) to every chunk at ingest, then filter the vector search by the requesting user's effective permissions before reranking. For high-sensitivity corpora we add per-tenant index isolation. SharePoint, Confluence, Google Drive, and Slack connectors all support this when configured correctly — misconfigured RAG is a common source of accidental data exposure, so we audit this explicitly.
Three tiers. RAG audit is 8,500 EUR over two weeks: corpus profile, current-state assessment if you have an existing RAG, architecture recommendation, and an eval-set design. RAG pilot is 35,000 EUR over 6 to 8 weeks for up to 10,000 documents: ingestion pipeline, vector store, hybrid retrieval, reranking, eval harness, and a customer-zero deployment. Production RAG retainer starts at 14,000 EUR per month and covers corpus expansion, embedding upgrades, eval growth, and on-call. Typical path from kickoff to production-grade RAG is 8 to 12 weeks.
