Clinical metagenomics is a broad-range agnostic detection method of pathogens, including novel microorganisms. A major limit is the low pathogen load compared to the high background of host nucleic acids. To overcome this issue, several solutions exist, such as applying a very high depth of sequencing, or performing a relative enrichment of viral genomes associated with capsids. At the end, the quantity of total nucleic acids is often below the concentrations recommended by the manufacturers of library kits, which necessitates to random amplify nucleic acids. Using a pool of 26 viruses representative of viral diversity, we observed a deep impact of the nature of sample (total nucleic acids versus RNA only), the reverse transcription, the random amplification and library construction method on virus recovery. We further optimized the two most promising methods and assessed their performance with fully characterized reference virus stocks. Good genome coverage and limit of detection lower than 100 or 1000 genome copies per mL of plasma, depending on the genome viral type, were obtained from a three million reads dataset. Our study reveals that optimized random amplification is a technique of choice when insufficient amounts of nucleic acid are available for direct libraries constructions.