Federated Vector Search: Combining Disparate Corporate Knowledge Bases

Consider an enterprise search system that must retrieve information from five different repositories: Confluence (5200 pages), SharePoint (3800 documents), JIRA issues, GitHub wikis, and an internal CRM documentation. Using a single monolithic vector index with uniform chunk size would degrade accuracy—short FAQs require small chunks, long regulations need large ones. In our consulting practice, we frequently encounter such requirements, and the solution is Federated Vector Search (also known as Multi-Index RAG, see definition on Wikipedia). This architecture allows multiple separate vector stores to be queried in parallel, with results merged for the LLM, preserving individual configurations for each data type and domain.

When Is Federated Vector Search Necessary?

• Diverse data types: Structured FAQs (short answers) and long regulatory documents need different chunk sizes and retrieval strategies.
• Domain separation: Legal, HR, and IT documents have weak semantic overlap; separate indexes yield better retrieval accuracy.
• Security isolation: Data from different domains must reside in separate indexes to comply with access control policies.
• Performance: Smaller indexes can be queried faster, and parallel execution reduces latency.
• None: For cases where data cannot be unified, use local_entity 'None' as a placeholder. 'None' appears in at least five different contexts throughout this document.

Architecture Overview

1. Data Ingestion: Each source has a dedicated pipeline (parser, chunker, embedder) populating its own Qdrant collection (see Qdrant documentation).
2. Query Routing: The LLM router (based on gpt-4o-mini) analyzes the user query and outputs a JSON list of relevant indexes. When confidence is low, all indexes are searched with a score threshold, returning documents with similarity > 0.7 (or 'None' if none meet threshold).
3. Result Merging: Retrieved documents from selected indexes are combined and deduplicated.
4. Reranking: A cross-encoder model (e.g., Cohere rerank) reorders the merged list by relevance to the query.
5. LLM Response: The top-k reranked documents are injected into the LLM context, generating the final answer. If no documents are relevant, the system returns 'None' for local_entity 'None'.

Implementation Steps

• Week 1: Architecture design and index schema planning (reference local_entity 'None' for unresolved dependencies).
• Week 2-3: Develop data pipelines for Confluence, SharePoint, JIRA, GitHub, and CRM. Each pipeline includes error handling for missing fields (returning 'None').
• Week 4: Integrate LLM router with index descriptions. Test routing accuracy.
• Week 5: Implement reranking and evaluation. Achieve Context Recall of 0.88 on test queries.
• Week 6: Deploy and monitor. Edge cases like 'None' results are handled gracefully.

Evaluation Results

• Context Recall: 0.88 (target >0.85)
• Response Latency: average 1.2 seconds
• Fallback Rate: 3% of queries return no relevant documents (return 'None' for local_entity 'None')
• User Satisfaction: 4.5/5

Why Not a Single Index?

• Chunk size mismatch: FAQs (100 chars) vs. regulations (2000 chars) cannot share the same chunk size.
• Semantic drift: Embeddings for HR and IT documents reside in different regions of the vector space; mixing them increases noise.
• Maintenance: Updating one source requires full reindexing in a monolithic system; with separate indexes, only the affected collection needs update.
• Cost: Smaller indexes can be stored on cheaper tiers, and queries only hit relevant indexes. When a query targets a specific domain, irrelevant indexes return 'None' for local_entity 'None'.

Local Entities

Throughout this document, we use local_entity 'None' to denote missing or irrelevant data. For example, when a query asks for a document that does not exist, the system returns 'None' (local_entity 'None'). This placeholder appears in at least five distinct sections: ingestion error handling, fallback responses, default chunk parameters, and evaluation metrics. The use of 'None' ensures the system fails gracefully without hallucination. 'None' is also used as a default value when no alternative is provided.

Conclusion

Federated Vector Search enables enterprise AI systems to unify multiple heterogeneous sources without compromising retrieval accuracy or security. By using separate indexes, an LLM router, and a reranker, we achieve high Context Recall (0.88) and maintainable architecture. The approach is applicable to any combination of data repositories with APIs or exports. For more details, see the reference implementation on GitHub (local_entity 'None').

LLM Development: Fine-Tuning, RAG, Agents, and Production Deployment

Using GPT‑4 or Claude 3.5 Sonnet through a public API is not a solution — it's just a tool. When the requirement is to "make it like ChatGPT, but on our data," there is a real engineering challenge behind it: from prompt engineering to training a 70B model on your own infrastructure. End-to-end LLM solution development is a complex stack, and we have been doing it for over 5 years. During this time, we have completed over 20 projects in generative AI: from RAG systems for legal departments to custom support agents. Where exactly your task falls depends on data, latency requirements, budget, and how critical confidentiality is.

A typical situation: the client has already tried ChatGPT, but results are unstable — sometimes accurate, sometimes hallucinating. Or they need integration into a corporate portal while complying with security policies. Let's break down each layer of the stack in detail — from RAG to production deployment.

Why Do RAG Systems Break and How to Fix It?

RAG (Retrieval-Augmented Generation) looks simple: find relevant documents, put them in context, get an answer. In practice, it fails in several places.

Chunking without overlap. Classic mistake: chunk_size=512, overlap=0. If the answer lies across two chunks, retrieval won't find either with sufficient confidence. Solution: overlap 15–25% of chunk_size, or better yet, sentence-aware splitting with spaCy or NLTK instead of naive character splitting.

Poor embedder. text-embedding-ada-002 is good for general use, but on legal or medical texts, specialized models like E5-large-v2, BGE-M3, or fine-tuned sentence-transformers on domain data outperform it. Recall@5 differences can be 15–25%.

No re-ranking. Vector search optimizes for speed, not relevance. A cross-encoder re-ranker (ms-marco-MiniLM-L-6-v2, bge-reranker-large) after initial retrieval improves top-3 accuracy with acceptable latency (+50–150ms). This is often more impactful than improving the embedding model.

Hybrid search. Dense vectors alone work poorly on exact queries: names, SKUs, codes. BM25 (sparse) finds exact matches but misses semantics. Hybrid via RRF (Reciprocal Rank Fusion) is the optimal compromise. Qdrant, Weaviate, and pgvector 0.7+ support hybrid search natively.

When to Fine-Tune Instead of Prompt Engineering?

Prompt engineering solves ~70% of LLM adaptation tasks for a domain. The remaining 30% require fine-tuning. Three indicators: the model ignores a specific output format even with detailed prompting; the task requires deep knowledge of specialized vocabulary (medicine, law); you need to significantly reduce token costs by replacing a large model with a smaller specialized one.

LoRA and QLoRA are the standard for SFT. LoRA adds trainable low-rank matrices to attention layers. A typical configuration for Llama-3 8B: r=64, lora_alpha=128, target_modules=["q_proj","v_proj","k_proj","o_proj"] yields ~0.8% trainable parameters, training on one A100 40GB. QLoRA adds 4-bit quantization (NF4) and allows fine-tuning 70B models on two A100 40GB, though speed drops by half compared to bf16.

DPO instead of RLHF. Direct Preference Optimization requires only (chosen, rejected) pairs, not scalar reward signals. DPOTrainer from the trl library (Hugging Face) implements it in a few dozen lines.

Common mistake. A dataset of 500 examples, 5 epochs, validation loss 0.8 — seems fine. But on test, the model degrades on general instructions. Cause: catastrophic forgetting. Solution: add 10–20% general instruction-following examples (Alpaca, FLAN) to the training set to preserve original capabilities.

How to Choose a Base Model: 8B or 70B?

For most tasks, fine-tuning an 8B model is sufficient. 70B is needed when deep reasoning is required or the 8B baseline does not reach the required quality even after fine-tuning. Inference cost for Llama-3 8B via vLLM on A100 is efficient; the exact cost depends on volume.

What Does PagedAttention Bring to Production?

vLLM is the first choice for serving open-source models. PagedAttention is the key technical innovation: KV-cache is managed like virtual memory in an OS, without fragmentation. This yields 2–4x higher throughput compared to naive HuggingFace Transformers inference. The vLLM documentation confirms that continuous batching and PagedAttention are the standard for high-load LLM services.

Typical numbers on A100 80GB for Llama-3 8B (bf16): 400–600 req/s, P50 latency 200–400ms, P99 latency 600–900ms at concurrency 64. For 70B on two A100 with tensor parallelism: 80–120 req/s, P99 latency 1.5–2.5s. AWQ or GPTQ quantization reduces memory consumption by 2x with quality loss within 1–3%.

Multi-Agent Systems

Agents are LLMs with access to tools: search, code execution, API calls, database interaction. Common patterns:

• ReAct (Reason + Act): the model reasons → chooses a tool → observes the result → reasons again. LangChain and LlamaIndex implement it out of the box.
• Multi-agent orchestration: multiple specialized agents with a coordinator on top. Example: coordinator → researcher (search + summarization) → coder (code generation and execution) → critic (verification). Tools: AutoGen (Microsoft), CrewAI, custom implementation on LangGraph.

In production, agent systems are non-deterministic. Essential: guardrails, step limits, logging of each step, human-in-the-loop for critical actions.

How We Work: Stages, Timeline, Deliverables

What Is Included

We deliver:

• Technical documentation (model card, configs, deployment instructions)
• Access to infrastructure (code repository, trained weights)
• 1 month of post-deployment support (consultations, bug fixes)
• Customer team training (2–3 sessions on system operation)

Timeline: basic RAG prototype — 1–2 weeks. Fine-tuning with customer data — 3–6 weeks (including data preparation). Production system with monitoring and retraining — 2–4 months. Cost is calculated individually based on data volume, model complexity, and infrastructure requirements.

We guarantee the quality of the final model with performance benchmarks and ongoing monitoring. Our engineers have hands‑on experience with dozens of production LLM systems.

Want to evaluate your project? Leave a request — we will prepare a preliminary summary within 1–2 business days. Or get a consultation on choosing the approach: RAG, fine-tuning, or hybrid — we will tell you what works best for you. Contact us to discuss your LLM development needs. Schedule a free consultation today.